Wind wave and flow combined generation power prediction method and system based on RF-GBRT

Abstract

The invention provides an RF-GBRT-based wind wave and flow combined power generation power prediction method and system. The method comprises the following steps: S1, obtaining a historical data set and real-time collection data; s2, constructing a fusion integrated prediction model; s3, preprocessing the data acquired in real time; s4, inputting the preprocessed real-time features, and generating a power generation power prediction value; and S5, performing closed-loop optimization on the model. Through multi-algorithm deep collaboration and dynamic optimization, the problem of power prediction in a wind wave and flow coupling scene is effectively solved, and technical support is provided for efficient operation of a composite power generation system.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and intelligent prediction technology, and in particular to a method and system for predicting combined wind, wave and flow power generation based on RF-GBRT. Background Technology

[0002] As the global energy structure shifts towards cleaner energy, marine renewable energy (wind, wave, and tidal current energy) has become a development hotspot due to its abundant reserves and strong stability. Wind-wave-tidal combined power generation devices can compensate for the intermittent nature of single-energy sources and improve energy efficiency by synergistically utilizing multiple marine energy sources. However, their power generation is affected by the coupling of multiple environmental parameters such as wind speed, wind direction, and tidal current velocity, and there are strong nonlinear relationships between these parameters (e.g., a sudden increase in wind speed may exacerbate power fluctuations due to changes in tidal current direction), making power prediction significantly more difficult than for single-energy power generation systems.

[0003] Existing prediction methods have the following limitations: 1. A single algorithm (such as using random forest or gradient boosting tree alone) is difficult to balance feature mining ability and nonlinear fitting accuracy at the same time, and is prone to prediction bias in complex coupled scenarios. 2. Traditional models do not consider the dynamic changes in the importance of environmental features (such as seasonal changes that may alter the weighting of wind speed and tidal flow velocity on power), resulting in a rapid decline in long-term prediction accuracy. 3. Lacking a closed-loop optimization mechanism, it is unable to dynamically adjust model parameters based on real-time operational data, making it difficult to adapt to the time-varying characteristics of the marine environment. Summary of the Invention

[0004] This invention provides a method and system for predicting combined wind, wave and current power generation based on RF-GBRT. Through a full-process design of data acquisition, feature interaction, model fusion, real-time prediction and closed-loop optimization, it achieves high-precision power prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The method for predicting combined wind, wave, and current power generation based on RF-GBRT includes the following steps: S1: Obtain historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical dataset, and collect current environmental parameters in real time as real-time dataset; S2: Construct an RF-XGBoost-GBRT fusion prediction model, including: using historical environmental parameters as input and historical power generation data as output, training a random forest model to obtain normalized feature importance weights and random forest model prediction values; weighting historical environmental parameters based on normalized feature importance weights to obtain extreme gradient boosting weighted features; using extreme gradient boosting weighted features as input and historical power generation data as output, training an extreme gradient boosting model, outputting preliminary extreme gradient boosting prediction values ​​and calculating residuals; using historical environmental parameters as input and residuals as labels, training a gradient boosting regression tree model, outputting gradient boosting regression tree corrected prediction values; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation, and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model; S3: Correct outliers and normalize the real-time dataset, then generate the weighted real-time features required for the extreme gradient boosting model based on the importance weights of the normalized features, and output the normalized real-time original features and weighted real-time features. S4: Input the normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input the weighted real-time features into the extreme gradient boosting model to obtain the real-time predicted values ​​of each model. Then, sum them according to the fusion weights to output the real-time power generation prediction value. S5: Combine the normalized real-time original features, the real-time power generation prediction value, and the corresponding actual real-time power generation to form a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integrated prediction model.

[0006] In this specification, step S2, the training process of the random forest model includes: constructing multiple decision trees, obtaining training samples through bootstrap sampling, randomly selecting some historical environment parameters when splitting each tree, splitting based on the node impurity minimization criterion, calculating the total reduction in node impurity for each type of historical environment parameter as the feature importance weight, and then performing normalization processing.

[0007] In this specification, in step S2, the weighted features input to the extreme gradient boosting model are the product of historical environmental parameters and the importance weights of normalized features. The predicted values ​​are iteratively optimized through the gradient boosting strategy, so that the predicted values ​​gradually approach the difference between historical power generation data and the predicted values ​​of the random forest model.

[0008] In this specification, in step S2, the interaction process between the gradient boosting regression tree model and the extreme gradient boosting model is as follows: the training label of the gradient boosting regression tree model is the residual, and the unfitting bias in the residual is learned iteratively through the gradient boosting strategy. The output gradient boosting regression tree corrected prediction value is used to compensate for the prediction bias of the random forest model and the extreme gradient boosting model.

[0009] In this specification, in step S2, the process of determining the fusion weights is as follows: using 5-fold cross-validation, the root mean square error of the three models on the validation set is calculated, the fusion weights are inversely proportional to the root mean square error, and the sum of the fusion weights of the three models is 1.

[0010] In this manual, in step S1, the historical environmental parameters are wind speed, wind direction, and tidal current velocity; the real-time environmental parameters are real-time wind speed, real-time wind direction, and real-time tidal current velocity.

[0011] In this specification, in step S3, the outlier correction adopts the 3σ principle to determine the abnormal real-time environmental parameters. The outlier is replaced by the mean of the parameters in the previous 3 normal times. In the 3σ principle, the mean and standard deviation are calculated based on the corresponding environmental parameters in the historical dataset.

[0012] In this specification, step S5, the closed-loop optimization process includes short-term optimization and long-term optimization; short-term optimization involves adjusting the learning rate and number of trees of the extreme gradient boosting model and the gradient boosting regression tree model every 1000 new samples accumulated; long-term optimization involves re-executing step 2 every 10000 new samples accumulated to update the normalized feature importance weights, parameters of each sub-model, and fusion weights.

[0013] In this specification, step S3, the normalization process includes: based on the minimum and maximum values ​​of each environmental parameter in the historical dataset, mapping the corrected real-time environmental parameters to the [0,1] interval to obtain normalized real-time features.

[0014] The RF-GBRT-based wind-wave-current combined power generation prediction system, applying any one of the above-mentioned RF-GBRT-based wind-wave-current combined power generation prediction methods, comprises: The data acquisition module is used to acquire historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical datasets, and to collect current environmental parameters in real time as real-time datasets. The model building module is used to construct the RF-XGBoost-GBRT fusion prediction model, including: training a random forest model to obtain normalized feature importance weights and random forest model prediction values ​​using historical environmental parameters as input and historical power generation data as output; weighting historical environmental parameters based on normalized feature importance weights to obtain weighted features; training an extreme gradient boosting model using weighted features as input and historical power generation data as output, outputting the initial prediction values ​​of extreme gradient boosting and calculating the residuals; training a gradient boosting regression tree model using historical environmental parameters as input and residuals as labels, outputting the corrected prediction values ​​of gradient boosting regression tree; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation; and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model. The real-time preprocessing module is used to correct outliers and normalize the real-time dataset. Then, based on the importance weights of the normalized features, it generates the weighted real-time features required for the extreme gradient boosting model and outputs the normalized real-time original features and weighted real-time features. The prediction module is used to input normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input weighted real-time features into the extreme gradient boosting model to obtain the real-time prediction values ​​of each model. Then, the prediction values ​​are weighted and summed according to the fusion weights to output the real-time power generation prediction value. The model optimization module is used to combine the normalized real-time original features, the real-time power generation prediction value and the corresponding actual real-time power generation into a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integrated prediction model.

[0015] In summary, the present invention has at least the following beneficial effects: 1. Improve prediction accuracy: Through the deep interaction of three algorithms (random forest to mine feature importance, extreme gradient boosting to focus on key features to output preliminary prediction, and gradient boosting regression tree to correct residuals), the coupling relationship and nonlinear characteristics of multiple environmental parameters are effectively captured, reducing prediction bias in complex scenarios. 2. Enhanced model robustness: The fusion algorithm integrates the overfitting ability of random forest, the regularization advantage of extreme gradient boosting, and the bias correction ability of gradient boosting regression tree, making it more adaptable to abnormal environmental data (such as sudden strong winds and reversed tides). 3. Dynamically adapt to environmental changes: The closed-loop optimization mechanism automatically adapts to the dynamic changes in the importance of environmental features (such as the impact of seasons on the wind speed-power relationship) through real-time data updates and model iterations, maintaining the stability of prediction accuracy in the long term; 4. Supporting efficient operation and scheduling: High-precision prediction results can provide a reliable basis for real-time control of power generation devices (such as blade angle adjustment and energy storage system coordination) and grid scheduling, thereby improving energy utilization efficiency and grid connection stability. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of the wind-wave-current combined power generation prediction method based on RF-GBRT involved in this invention.

[0018] Figure 2 This is a schematic diagram illustrating the process of constructing the fusion prediction model involved in this invention.

[0019] Figure 3 This is a schematic diagram of the real-time data preprocessing process involved in this invention.

[0020] Figure 4 This is a schematic diagram of the real-time power prediction process involved in this invention. Detailed Implementation

[0021] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0022] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this embodiment provides a method for predicting combined wind, wave, and current power generation based on RF-GBRT, including the following steps: S1: Obtain historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical dataset, and collect current environmental parameters in real time as real-time dataset; S2: Construct an RF-XGBoost-GBRT fusion prediction model, including: using historical environmental parameters as input and historical power generation data as output, training a random forest model to obtain normalized feature importance weights and random forest model prediction values; weighting historical environmental parameters based on normalized feature importance weights to obtain extreme gradient boosting weighted features; using extreme gradient boosting weighted features as input and historical power generation data as output, training an extreme gradient boosting model, outputting preliminary extreme gradient boosting prediction values ​​and calculating residuals; using historical environmental parameters as input and residuals as labels, training a gradient boosting regression tree model, outputting gradient boosting regression tree corrected prediction values; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation, and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model; S3: Correct outliers and normalize the real-time dataset, then generate the weighted real-time features required for the extreme gradient boosting model based on the importance weights of the normalized features, and output the normalized real-time original features and weighted real-time features. S4: Input the normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input the weighted real-time features into the extreme gradient boosting model to obtain the real-time predicted values ​​of each model. Then, sum them according to the fusion weights to output the real-time power generation prediction value. S5: Combine the normalized real-time original features, the real-time power generation prediction value, and the corresponding actual real-time power generation to form a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integration model.

[0025] The technical approach of this invention is as follows: Based on historical operating data (including environmental parameters such as wind speed, wind direction, and tidal current velocity, and their corresponding actual power generation) and real-time collected data from the wind-wave-current combined power generation device, the scheme achieves high-precision prediction through a three-level algorithm collaboration and closed-loop optimization: First, Random Forest (RF) is used to mine the influence weights of environmental parameters on power generation, providing a feature-focusing basis for Extreme Gradient Boosting (XGBoost); XGBoost outputs preliminary predicted values ​​based on weighted features and passes the unfitted residuals to Gradient Boosting Regression Tree (GBRT); GBRT specifically corrects the residuals, and the three are then fused through dynamic weighting to form the final prediction model. Simultaneously, historical datasets are updated through real-time data feedback, and model parameters and weights are periodically iterated and optimized to adapt to dynamic environmental changes. Specific details are as follows: S1: Acquire two types of core data (historical datasets and real-time collected data). This step provides foundational data support for subsequent model training and prediction, and it is essential to ensure the data's integrity, temporal consistency, and physical correlation. The data content covers key environmental parameters affecting the combined power generation of wind, waves, and current, along with corresponding actual power records, as detailed below: 1.1 Data Content and Symbol Definitions 1. Historical dataset: Contains environmental parameters and actual power generation matching samples accumulated during device operation, used for training and validation of the fusion model.

[0026] Environmental parameters (input features): enumerated into 3 types of core parameters, namely wind speed (denoted as...). (Unit: m / s), wind direction (denoted as...) Unit: °, with true north as 0° and increasing clockwise), tidal current velocity (denoted as ). (Unit: m / s)

[0027] Actual power generation (output label): denoted as y (unit: kW), which is the actual output power of the device under the corresponding environmental parameters.

[0028] Sample format: Historical datasets can be represented as , where n is the total number of samples (enumerated values: at least 10,000 to ensure coverage of different seasons and weather conditions). The environmental feature vector of the i-th sample ( This represents the value of the j-th type of environmental parameter for the i-th sample (j=1,2,3). Let be the actual power generation of the i-th sample.

[0029] 2. Real-time data acquisition: Used to generate real-time power generation prediction values, with parameter types completely consistent with historical datasets.

[0030] Real-time environmental parameters: denoted as real-time wind speed Real-time wind direction Real-time current flow rate The format of a single data collection is .

[0031] 1.2 Data Sources and Collection Requirements Historical dataset source: Environmental parameters: These parameters are obtained from three types of sensors equipped with the device. Wind speed is collected by an ultrasonic anemometer (accuracy ±0.1m / s), wind direction is collected by a mechanical wind vane (accuracy ±1°), and tidal current velocity is collected by an electromagnetic current meter (accuracy ±0.01m / s). Actual power generation: Power metering device from the grid connection point (accuracy ±0.5%); Time consistency requirement: The timestamps of all parameters must be precisely aligned (error ≤ 1 second) to ensure the physical correlation between environmental parameters and power samples (e.g., wind speed, wind direction, and flow velocity at 10:00:00 must correspond to the actual power at 10:00:00).

[0032] Real-time data acquisition source: The same sensors and measuring devices are used as the historical dataset, and the acquisition frequency is enumerated at 1 time / minute (balancing real-time performance and data storage efficiency). The data is transmitted to the prediction system in real time via industrial bus.

[0033] 1.3 Data Preprocessing (Preliminary Cleaning) To reduce the impact of outlier data on subsequent model training, the historical dataset was initially cleaned: Remove samples with missing or inconsistent timestamps (such as samples that only record wind speed but lack the corresponding power). Remove outliers that are clearly outside the physical range (such as unreasonable values ​​like wind speed > 70 m / s or tidal current velocity < 0 m / s). Missing values ​​(representing less than 5%) are filled using linear interpolation. For example, if the wind speed of the i-th sample is missing, it is filled using linear interpolation. Calculation (based on normal data from adjacent time points).

[0034] Output: Historical dataset after initial cleaning and real-time data stream acquisition This provides input for model training in S2 and real-time preprocessing in S3.

[0035] S2: Constructing an RF-XGBoost-GBRT fusion prediction model (training the three algorithms collaboratively) This step is the core fusion process, achieving deep interaction between three algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Gradient Boosting Regression Tree (GBRT) through a feature importance transfer, residual progressive correction, and dynamic weight fusion mechanism. Specifically, RF is responsible for mining the influence weights of environmental features on power generation, providing XGBoost with input focusing on key features; XGBoost outputs preliminary predictions based on weighted features and passes the unfitted residuals to GBRT; GBRT performs fine-tuning on the residuals, and finally, the prediction results from the three algorithms are fused through dynamic weights to form a model that combines generalization ability and prediction accuracy. The input data is the historical dataset in S1. (n is the total number of samples, Let i be the environmental feature vector of the i-th sample, including wind speed. ,wind direction tidal current velocity ; (where is the actual power generation of the i-th sample), and the output is the trained fusion model. The process of building the fusion prediction model is as follows: Figure 2 As shown.

[0036] 2.1 Random Forest (RF) Model Construction and Training: Output Feature Importance Weights The RF model reduces the risk of overfitting by integrating multiple independent decision trees. Its core function is to quantify the impact of various environmental features on power generation and output the feature importance weights, providing a basis for feature focusing in the subsequent XGBoost model.

[0037] 2.1.1 Model Construction Details RF model by The decision tree consists of ( ) To preset the hyperparameters, the enumerated values ​​are 100, 200, and 300. The optimal value needs to be selected through a validation set. This scheme uses 100 as an example. The construction of each tree follows a double random principle: Randomized samples: Sampling from historical datasets via bootstrap n samples are drawn from the sample (sampling with replacement, repeated samples are allowed), which are used as the nth sample. Trees ( Training samples; Randomized features: When a tree splits, it is randomly selected from m=3 environmental features (wind speed, wind direction, and tidal current velocity). Features ( To preset hyperparameters, the value is usually taken as follows: This means randomly selecting two features each time for node splitting, to avoid a single feature dominating the structure of the entire tree.

[0038] Each decision tree is a binary tree, with leaf nodes outputting the mean power generation of the corresponding sample subset (a characteristic of regression trees). The depth of the tree is controlled by pre-pruning (a hyperparameter). The enumeration values ​​are 5, 8, and 10 to prevent the tree from becoming too deep and causing overfitting.

[0039] 2.1.2 Model Training Process 1. Single tree training: For the first tree... A tree is constructed, with environmental characteristics X from bootstrap sampling as input and actual power generation y as output. Nodes are split using a node impurity minimization criterion. Node impurity is measured using mean squared error (MSE), calculated as follows: ; in, Indicates the node currently to be split. This represents the number of samples contained in that node. This represents the average actual power generation of the sample at that node. During splitting, the power generation is determined from a randomly selected node. For each feature, iterate through all possible splitting thresholds for each feature, select the feature and threshold that maximizes the difference between the impurity of the parent node and the sum of the impurities of the child nodes, and complete the node splitting.

[0040] 2. Feature Importance Calculation: After all trees are trained, the importance of each environmental feature is calculated—measured by the total reduction in node impurity caused by that feature across all trees. For the j-th feature (j=1,2,3, corresponding to wind speed, wind direction, and tidal current velocity respectively), its RF feature importance weight is calculated. The calculation formula is: ; in, Indicates the first All splits performed on the j-th feature in the trees; The impurity of the parent node of this split; This represents the impurity of the two child nodes after the split. For example: if the wind speed (j=1) cumulatively reduces the node impurity by 500 across 100 trees, the wind direction (j=2) by 300, and the tidal current speed (j=3) by 200, then... =500, =300, =200.

[0041] 3. Weight Normalization: To ensure the sum of feature weights is 1 (for easier subsequent weighted calculations), the weights are normalized. Normalization is performed to obtain the normalized feature importance weights. : ; Following the previous example, after normalization =500 / (500+300+200)=0.5, =0.3, =0.2, indicating that wind speed has the highest weight in terms of power generation.

[0042] 2.1.3 Model Output and Interaction Connection The RF model outputs two key results: Normalized feature importance weights It is directly used as the feature weighting coefficient of the XGBoost model (see 2.2.1) to realize the feature focusing interaction between RF and XGBoost; RF model's predicted values ​​for samples (The mean of all decision tree predictions), calculated using the following formula: ; in, For the first The prediction value of the decision tree for the i-th sample (i.e., the mean power of the leaf node containing the sample). This prediction value will be used in subsequent residual calculations (see 2.2.3) to provide a basis for correcting the GBRT model.

[0043] 2.2 XGBoost Model Construction and Training: Based on RF weighted feature input, the model outputs preliminary predicted values ​​and residuals. The XGBoost model optimizes prediction accuracy through a gradient boosting strategy. Its core function is to use the feature weights output by RF to focus on key environmental parameters, output more accurate preliminary predictions, and pass the unfitted residuals to GBRT to achieve residual correction interaction between XGBoost and GBRT.

[0044] 2.2.1 Model Construction Details The XGBoost model is based on Composed of gradient boosting trees ( To preset the hyperparameters (enumerating values ​​of 150, 200, and 250; this scheme uses 150 as an example), a regularization term is introduced to control the tree complexity (to avoid overfitting). The objective function is: ; in, This is the set of parameters for the XGBoost model; The loss function (using mean squared error) For regularization terms ( For the first The number of leaf nodes in a tree. The output weight of the l-th leaf node is... The regularization coefficient is enumerated, and its values ​​are... , ).

[0045] Unlike traditional XGBoost, the input features of XGBoost in this scheme are RF-weighted environmental features, with the aim of enhancing the influence of important features.

[0046] 2.2.2 Model Training Process 1. Feature weighting: The original environmental feature matrix X (each row corresponds to one sample, and each column corresponds to one feature) from the historical dataset is weighted by the normalized weights of the RF output. Multiplying them together yields the input feature matrix of XGBoost. The weighted feature vector of the i-th sample is: ; Following the example in 2.1.2, if the original features of a sample are... Then after weighting It can be seen that the weighted value of wind speed has the highest proportion, which is consistent with the importance of RF identification.

[0047] 2. Initializing Predictions: XGBoost iteratively optimizes predictions, starting with the initial prediction values. Take the average of the actual power generation of all samples (simplifying the initial state): .

[0048] 3. Iterative training of gradient boosting trees: For each tree Perform the following steps: Calculate the current residual: the i-th sample at the i-th position... The predicted value of the wheel Then the residual (i.e., the negative gradient of the loss function) is: ; Training Trees: by weighted features For input, residual For output, train the regression tree (The tree structure is determined by optimizing the objective function); Update predictions: Introduce learning rate (Preset hyperparameters, enumerated values ​​of 0.1 and 0.05, to control the contribution of each tree and avoid overfitting), update predicted values: .

[0049] 2.2.3 Model Output and Interaction Connection The XGBoost model outputs two key results: Final preliminary forecast (Predicted value after 150 iterations); residual (That is, actual power - RF prediction - XGBoost prediction, used as the training label for GBRT), the calculation formula is: ; The residual reflects the misfitting bias of RF and XGBoost (such as nonlinear error when wind, waves and current are strongly coupled). GBRT will learn these biases in a targeted manner to achieve the correction interaction between XGBoost and GBRT.

[0050] 2.3 GBRT Model Construction and Training: Based on XGBoost residual input, output corrected predicted values The GBRT model focuses on residual correction through a gradient boosting strategy. Its core function is to learn the misfitting bias of RF and XGBoost, output corrected predictions, and make up for the lack of accuracy of the first two models in complex coupled scenarios.

[0051] 2.3.1 Model Construction Details The GBRT model is based on Composed of gradient boosting trees ( The hyperparameters are preset, with enumerated values ​​of 120, 150, and 180 (this scheme uses 120 as an example). The difference from XGBoost is that the input feature is the original environment feature X (no weighting required, retaining complete feature information), and the input label is the residual output by XGBoost. (Focusing on bias correction). Its objective function is: ; in, This is the set of parameters for the GBRT model; The loss function (using mean squared error) ; For regularization terms (the form is consistent with XGBoost, and the regularization coefficients are enumerated with values...), , (More lenient to accommodate the volatility of residuals).

[0052] 2.3.2 Model Training Process 1. Initialize residual predictions: Initial predictions for GBRT Take all residuals Mean: ; 2. Iterative training of gradient boosting trees: For each tree... Perform the following steps: Calculate the residuals (i.e., the corrected bias): the residual of the i-th sample in the i-th case... The predicted residual value of the wheel is The correction deviation is then: ; Training Trees: Based on the characteristics of the original environment For input and correction of deviation For output, train the regression tree ; Update residual predictions: Introduce learning rate (Preset hyperparameters, enumerated values ​​0.08 and 0.06; due to large residual fluctuations, the learning rate is slightly lower than XGBoost). Update predicted values: .

[0053] 2.3.3 Model Output and Interaction Connection The output of the GBRT model is the final residual correction prediction value. (The residual prediction value after 120 iterations). This value will be fused with the predictions from RF and XGBoost (see 2.4), and its physical meaning is to compensate for the bias of the first two models—for example, when wind speed increases suddenly and the tidal current direction changes abruptly, RF and XGBoost may underestimate the power, while GBRT's... It will output a positive value to compensate for prediction bias.

[0054] 2.4 Weight fusion of three algorithms: Outputting the final fused prediction model By dynamically weighting the predictions from RF, XGBoost, and GBRT, the advantages of each model are balanced (RF has strong generalization ability, XGBoost has accurate initial predictions, and GBRT has fine corrections), and the final prediction model is output.

[0055] 2.4.1 Determining Fusion Weights: Based on 5-fold Cross-Validation The prediction errors of the three models were calculated using 5-fold cross-validation, and weights were assigned inversely proportional to the errors (the smaller the error, the larger the weight). The steps are as follows: 1. Dataset partitioning: Randomly divide the historical dataset into 5 mutually exclusive subsets. (The number of samples in each subset is approximately n / 5); 2. Cross-validation iteration: For k=1,2,3,4,5, ... The validation set is used as the union of the remaining four subsets, which forms the training set. RF, XGBoost, and GBRT models are trained on these sets respectively, and the root mean square error (RMSE) on the validation set is calculated. RF's RMSE (Verification Set Requirement): ; XGBoost Validation Set RMSE: ; GBRT's RMSE (Verification Set of Validation) ; in, Let be the number of samples in the k-th validation set.

[0056] 3. Average Error Calculation: The average RMSE of the five validations is taken to obtain the overall error of each model. ; 4. Weight Calculation: Let the fusion weight of RF be... XGBoost is GBRT is The calculation formula is: ; satisfy ,and For example: If =8, =5, =10, then =0.125, =0.2, =0.1, the sum is 0.425, therefore =0.125 / 0.425≈0.294, =0.2 / 0.425≈0.471, =0.1 / 0.425≈0.235, XGBoost gets the highest weight because of the smallest error.

[0057] 2.4.2 Final Fusion Prediction Model Expression For any sample (historical sample or real-time sample), its predicted power generation value The weighted sum of the predictions from the three models is expressed as follows: ; The predicted value of the sample by the RF model (calculated based on the original feature X); The XGBoost model's prediction of this sample (based on RF weighted features). calculate); The GBRT model's prediction of this sample (calculated based on the original feature X); : Fusion weights determined through 5-fold cross-validation (fixed values, until model update).

[0058] 2.4.3 Model Performance Verification and Output The accuracy of the fusion model was evaluated using an independent validation set of the historical dataset (split from the original dataset in a 7:3 ratio, with 30% of the samples used as the validation set), and the mean absolute percentage error (MAPE) on the validation set was calculated. ; in, This is the number of samples in the validation set. If... If the preset accuracy standard is met, the model training is complete; otherwise, adjust the hyperparameters of the three algorithms (e.g., ...). Repeat steps 2.1-2.4 to retrain until the target is met.

[0059] Final output: The trained RF-XGBoost-GBRT fusion prediction model (including all parameters of the RF, XGBoost, and GBRT sub-models, and the fusion weights). This model will be used for subsequent real-time power generation prediction (see S4).

[0060] S3: Preprocess the real-time acquired data to match the input requirements of the fusion model. This step requires converting the real-time collected environmental parameters into feature data that meets the input format of the fusion model. The processing procedure must be consistent with the feature processing logic during model training in S2 to ensure prediction accuracy. The input data is the real-time collected data from S1. The normalized feature importance weights of the RF model output in S2 The output is real-time feature data adapted to each sub-model. The real-time data preprocessing workflow is as follows: Figure 3 As shown.

[0061] 3.1 Outlier Detection and Correction Real-time data may produce outliers due to momentary sensor malfunctions (such as wind speed suddenly dropping to 0 or far exceeding historical maximum values), which need to be handled through the following steps: 1. Outlier identification: The 3σ principle is adopted, and the mean value of each type of environmental parameter is calculated based on historical datasets. and standard deviation (j=1,2,3), if real-time parameters satisfy If the value is abnormal, it is considered an outlier. For example: historical wind speed data. =8m / s, =3m / s. If the real-time wind speed is 20m / s (20-8=12>3×3=9), then it is considered abnormal.

[0062] 2. Outlier Correction: Outliers are replaced using a sliding window mean method. The mean of the three previous normal values ​​for the parameter (if any) is calculated using the following formula: ; in, These are the normal values ​​of the j-th type of parameter at the first k time steps (k=1,2,3). These are the corrected real-time parameter values. If there are anomalies in the first 3 time points, the average of the 5 most recent normal time points is taken to ensure data continuity.

[0063] 3.2 Data Normalization Processing To eliminate the dimensional differences of different environmental parameters (such as wind speed in m / s and wind direction in °), the corrected real-time data needs to be normalized to the same range as when the S2 model was trained (usually [0,1]), as shown in the formula: ; in, , These are the minimum and maximum values ​​of the j-th type parameter in the S1 historical dataset (e.g., wind speed). =0m / s, =30m / s), These are the normalized real-time parameter values. The normalized data vector is denoted as... , used as input for the RF and GBRT sub-models (consistent with the feature scale during model training in S2).

[0064] 3.3 Generating XGBoost Weighted Real-Time Features Based on the feature importance weights output by the RF model in S2, the normalized real-time features are weighted to enhance the influence of key parameters. The weighting formula remains consistent with that in S2.2.2. ; For example: if The real-time features after normalization are Then after weighting It perfectly matches the training input format of XGBoost in S2.

[0065] Output: Normalized real-time raw features (For RF and GBRT sub-models), XGBoost weighted real-time features (For XGBoost sub-models), providing input for real-time predictions in S4.

[0066] S4: Input the preprocessed real-time features to generate predicted power generation values. This step, based on the fusion model trained in S2 and the preprocessed real-time features in S3, outputs the final predicted value through the collaborative computation of each sub-model. The core is to reproduce the fusion logic in S2 and apply it to real-time data. The input data is from S3. , and the S2 fusion model (including sub-model parameters and weights) The output is a real-time power generation prediction value. The real-time power prediction process is as follows: Figure 4 As shown.

[0067] 4.1 Real-time prediction calculation of each sub-model 1. RF sub-model prediction: The RF sub-model trained by S2 is input, and each decision tree outputs the average power of its leaf nodes based on the real-time feature paths. The final prediction is the average of all trees. ; in, For the first The prediction values ​​of decision trees for real-time features. =100 (consistent with S2).

[0068] 2. XGBoost sub-model prediction: Input the XGBoost sub-model trained by S2, and calculate the predicted value based on the iterative results of 150 gradient boosting trees: ; in, These are the initial predictions for XGBoost (from the S2 training results). =0.1 (learning rate, consistent with S2). For the first The predicted values ​​of the tree for weighted real-time features.

[0069] 3. GBRT sub-model prediction: Input the GBRT sub-model trained by S2, and calculate the residual correction value based on the iteration results of 120 gradient boosting trees: ; in, These are the initial residual predictions for GBRT (from S2 training results). =0.08 (learning rate, consistent with S2). For the first The predicted values ​​of each tree for real-time features.

[0070] 4.2 Fusion Calculation of Real-Time Predicted Values The real-time predictions of the three sub-models are weighted and summed according to the fusion weights \alpha, \beta, \gamma determined by S2 to obtain the final prediction: ; For example: if =0.294, =0.471, =0.235 (from example S2.4.1), the real-time predicted values ​​of the three sub-models are 120kW, 115kW, and 10kW respectively, then .

[0071] Output: Real-time power generation prediction of the wind-wave-current combined power generation device This value can be used in two types of scenarios: Device control: such as adjusting the blade angle of attack based on the predicted power (reducing the angle of attack when the predicted power is too high to avoid overload); Grid dispatch: Provide short-term power references for the grid and optimize the charging and discharging strategies of energy storage systems (such as initiating energy storage discharge in advance when power is predicted to be insufficient).

[0072] S5: Model Closed-Loop Optimization (Integration of Data Update and Iterative Optimization) To address the dynamic changes in the relationship between environmental characteristics and power generation (such as seasonal shifts in the wind speed-power curve), closed-loop optimization is needed to continuously improve model accuracy. Input data includes historical datasets for S1 and predicted values ​​for S4. Actual real-time power of the device ( The power source is the power metering device at the grid connection point of the wind-wave-current combined power generation device, which is consistent with the historical power generation acquisition device in S1. This power is the device's current real-time environmental parameter, i.e., the power collected by S1. The actual output power (and precisely aligned with the timestamps of real-time acquired data and real-time predicted values) and the preprocessing features of S3 are used to output the updated dataset and the optimized fusion model.

[0073] 5.1 Historical Dataset Update New samples are created by combining real-time collected features, predicted values, and actual values, and then added to the historical dataset in timestamp order to ensure data timeliness. New sample features: using the corrected original real-time parameters (Retain physical dimensions for easy renormalization later) New sample label: Actual real-time power ; Auxiliary information: Record predicted values It is used to analyze and predict error trends.

[0074] The updated historical dataset is , where k is the number of new samples (60 new samples per hour, based on a collection frequency of 1 time / minute).

[0075] 5.2 Iterative Optimization of the Fusion Model Based on the updated historical dataset, the S2 model training process is re-executed to achieve the following optimization objectives: 1. Feature Importance Weight Update: Retrain the RF model and calculate new feature importance weights. The influence of environmental parameters on weight changes (e.g., the impact of tidal current velocity on power may increase in winter, thus affecting the weight) (Increase).

[0076] 2. Sub-model parameter tuning: Retrain XGBoost and GBRT, and select better hyperparameters through the validation set (e.g., increase the number of trees in XGBoost when summer wind speed fluctuates greatly). Up to 200, improving fitting ability.

[0077] 3. Dynamic adaptation of fusion weights: Recalculate using 5-fold cross-validation. If GBRT's correction effect is more significant in the new data ( (reduced), then The weighting will be increased accordingly.

[0078] 5.3 Optimize frequency and triggering conditions To balance optimization cost and model accuracy, a hierarchical optimization strategy is adopted: Short-term optimization: Triggered every 1000 new samples accumulated (approximately 16.7 hours), only adjusting the learning rates of XGBoost and GBRT. With respect to the number of trees T, the RF feature weights and fusion weights remain unchanged (suitable for short-term environmental fluctuations). Long-term optimization: Triggered every 10,000 new samples (approximately 7 days), S2 (including RF feature weight calculation, sub-model training, and fusion weight update) is fully re-executed, suitable for long-term changes such as seasonal changes; Emergency optimization: When MAPE > 8% for 10 consecutive predicted values ​​(exceeding the preset threshold of 5%), full optimization is immediately triggered to resolve model mismatch caused by sudden environmental changes (such as abnormal wind speeds before a typhoon passes).

[0079] Output: The updated historical dataset and the optimized RF-XGBoost-GBRT fusion model are used for the next round of real-time prediction, forming a complete closed loop of data collection → model training → real-time prediction → data update → model optimization, ensuring long-term stability of prediction accuracy.

[0080] Through the above steps, this scheme achieves deep integration and dynamic optimization of the three algorithms. It utilizes the feature mining capabilities of RF and the initial prediction accuracy of XGBoost, and corrects the deviations in complex coupled scenarios through GBRT, ultimately providing high-precision and robust power prediction support for wind-wave-current hybrid power generation devices.

[0081] In some embodiments, S2 introduces a Multilayer Perceptron (MLP) to achieve deep interaction among four algorithms through a feature enhancement-residual iteration-deep fusion mechanism: RF mines feature importance, XGBoost outputs preliminary predictions based on weighted features, GBRT corrects the XGBoost residuals, and MLP integrates the intermediate results of the first three models with the original features to achieve nonlinear mapping. Finally, the predicted value is output through dynamic weight fusion. The input data is the historical dataset in S1. ( For environmental feature vectors, (This represents the actual power generation), and the output is the trained four-algorithm fusion model.

[0082] 1. Random Forest (RF) Model Construction and Training: Output Feature Importance Weights Model building and training (same as the original steps, with the addition of foreshadowing of interaction with MLP) RF model by =Composed of 100 decision trees, trained through bootstrap sampling and random feature selection, outputting normalized feature importance weights. This weight is not only used for feature weighting in XGBoost, but also serves as one of the input features of MLP, realizing the transfer of feature importance from RF to MLP.

[0083] Output Normalized feature importance weights ; RF Predictions (Used for subsequent residual calculations and MLP input).

[0084] 2. XGBoost Model Construction and Training: Based on RF weighted features, outputting preliminary predictions and residuals. Model building and training (same as the original steps, with the addition of interaction with MLP). XGBoost is powered by =Composed of 150 gradient boosting trees, with RF weighted features as input. The training process incorporates intermediate outputs of the MLP as additional regularization terms (to enhance the association with the MLP). Output: Preliminary forecast ; First residual (Passed to GBRT and MLP).

[0085] 3. GBRT Model Construction and Training: Correcting XGBoost Residuals and Outputting Quadratic Residuals Model building and training (same as the original steps, with the addition of interaction with MLP). GBRT is from =Composed of 120 gradient boosting trees, the input is the original features. The label is "first residual". The training process receives residual predictions from the MLP as a reference. Output: Residual correction value ; Second residual (Used as training labels for MLP to achieve residual transfer from GBRT to MLP).

[0086] 4. Multilayer Perceptron (MLP) Model Construction and Training: Deep Integration of Results from the First Three Models MLP integrates intermediate results from RF, XGBoost, and GBRT with original features through nonlinear mapping, making up for the insufficient fitting of tree models to high-dimensional coupled features. Its core function is to learn complex nonlinear relationships in strongly coupled wind, wave, and current scenarios.

[0087] Model building details MLP is a 3-layer neural network (input layer - hidden layer - output layer): Input layer: Dimension d=7, containing 7 types of features (integrating intermediate results from the first three models with the original features): ; in, Features of the original environment (3D). RF feature weights (3-dimensional). These are the predicted values ​​from the first three models (1 dimension), totaling 3+3+1=7 dimensions.

[0088] Hidden layer: number of neurons =16 (enumerated values ​​16 and 32, filtered through the validation set), the activation function used is ReLU (to solve nonlinear problems): ; in, The hidden layer weight matrix is... This is the hidden layer bias vector. This is the activation function.

[0089] Output layer: Number of neurons =1, output the correction value for the second residual. No activation function (regression task): ; in, This is the output layer weight matrix. This is the output layer bias.

[0090] Model training process 1. Loss function definition: using the second residual For the labels, mean squared error (MSE) loss is used, and a regularization term is introduced from the top three model predictions (to enhance synergy with other models): ; in, =0.01 (enumerated values ​​0.01 and 0.05) is the regularization coefficient, which constrains the deviation between the fusion result of the four models and the actual value.

[0091] 2. Parameter optimization: The Adam optimizer (learning rate) is used. =0.001, enumerating values ​​0.001, 0.005) minimize the loss function and update the parameters: ; ; 3. Early stopping strategy: Stop training when the validation set loss does not decrease for 10 consecutive rounds to avoid overfitting.

[0092] Output results and interactive connection MLP output residual final correction value This value will participate in the fusion of the four models, and at the same time, it will influence the training of XGBoost and GBRT in reverse through the regularization term of the loss function (achieving feedback interaction between MLP and XGBoost / GBRT).

[0093] 5. Weight fusion of four algorithms: Outputting the final fused prediction model 5.1 Determining Fusion Weights: Based on 5-fold Cross-Validation Extension Based on the original three-model cross-validation, MLP is added to calculate the validation set RMSE of the four models: ; in, .

[0094] Fusion weights (Corresponding to RF, XGBoost, GBRT, and MLP respectively) Distributed inversely proportional to the error, satisfying... : ; Example: If the RMSEs of the four models are 8, 5, 10, and 6 respectively, then The values ​​are 0.125, 0.2, 0.1, and 0.167 respectively, with a total of 0.592. Therefore... ≈0.211, ≈0.338, ≈0.169, ≈0.282.

[0095] 5.2 Final Fusion Prediction Model Expression The final predicted value is obtained by weighted sum of the predictions from the four models. : ; 5.3 Model Performance Validation Evaluation using the validation set MAPE is required. (The accuracy standard is improved due to the addition of MLP). If the standard is not met, the hyperparameters (such as the number of neurons in the MLP hidden layer and the number of XGBoost trees) are adjusted and the training is repeated.

[0096] Output: The trained RF-XGBoost-GBRT-MLP fusion model (including four model parameters and weights) ).

[0097] Key Interaction Relationships Description 1. RF and XGBoost / MLP: RF Simultaneously serving as the feature weighting coefficients for XGBoost and the input features for MLP; 2. XGBoost and GBRT / MLP: XGBoost's Participate in the calculation of the first residual This is passed to GBRT and MLP; 3. GBRT and MLP: GBRT's Participate in the calculation of the second residual , as training labels for MLP; 4. MLP and the top three models: The output of MLP affects the final result through the fusion weights, and its loss function regularization term inversely constrains the prediction bias of the top three models.

[0098] Through the above interactions, MLP makes up for the shortcomings of tree models in fitting nonlinear coupling features. The four models work together to improve prediction accuracy, especially in scenarios with strong coupling of wind, waves and current (such as typhoons accompanied by strong tidal currents).

[0099] Subsequent steps S3-S5 need to be adjusted accordingly, as detailed below: S3: Preprocess the real-time acquired data to match the input requirements of the four-algorithm fusion model. Additional input feature processing is needed for the MLP model to ensure that real-time data simultaneously meets the input formats of the four sub-models: RF, XGBoost, GBRT, and MLP. The input data is the real-time data acquired by S1. Normalized feature importance weights in the RF model of S2 The output is real-time feature data adapted to the four models.

[0100] Basic preprocessing (using the original logic, supplemented with MLP-related preparations) 1. Constant Value Detection and Correction: Outliers are handled using the 3σ principle, and the corrected real-time parameters are recorded as follows. (Formula as before).

[0101] 2. Data Normalization: Normalize the corrected data to the [0,1] interval to obtain... (Formula as above), used for raw feature inputs of RF, GBRT and MLP.

[0102] Sub-model dedicated feature processing 1. XGBoost weighted features: Same logic as S2, generated based on RF weights. (Formula as before).

[0103] 2. MLP Input Feature Construction: The MLP input is a 7-dimensional vector (integrating the original features, RF weights, and real-time predictions from the first three models). However, since the real-time predictions from the first three models need to be calculated in S4, the fixed part is prepared here first, and the dynamic part is supplemented in S4: Fixed component: Original normalized features (3D), RF feature weights (3D), totaling 6D; Dynamic part: Real-time predictions of RF, XGBoost, and GBRT (1-dimensional, to be calculated and supplemented in S4).

[0104] Output result: Normalized original features (For RF, GBRT, MLP); XGBoost Weighted Features (For XGBoost); MLP input fixed part (6-dimensional, dynamic part to be added in S4).

[0105] S4: Input the preprocessed real-time features to generate predicted power generation values. This step requires adding real-time prediction computation to the MLP, outputting the final predicted value through the collaborative output of four models. The input data consists of the preprocessed features from S3 and the four-algorithm fusion model from S2 (including sub-model parameters and weights). The output is the real-time predicted power generation value.

[0106] Real-time predictions from the first three sub-models (calculation of base prediction values) 1. RF sub-model: Input Output real-time predicted values: ; =100, For the first Decision tree prediction function; 2. XGBoost sub-model: Input Output real-time predicted values: ; =150, =0.1, These are the initial predicted values; 3. GBRT sub-model: Input Output real-time residual correction values: ; =120, =0.08, This represents the initial residual prediction value.

[0107] MLP sub-model real-time prediction (deep integration and correction) 1. Constructing the complete input vector: Integrate the fixed input (6-dimensional) of S3's MLP with the real-time predictions (1-dimensional) of the first three models to obtain a 7-dimensional input: .

[0108] 2. MLP forward computation: Hidden layer output (ReLU activated): ; For hidden layer weights, For bias, .

[0109] Output layer output (final residual correction value): ; For output layer weights, For bias.

[0110] The final predicted value is calculated by fusing four models. The weights of the four models determined based on S2 The weighted summation yields the real-time power generation prediction value: ; Example: If =0.211, =0.338, =0.169, =0.282, the predicted values ​​of the four sub-models are 120kW, 115kW, 10kW, and 5kW respectively, then: .

[0111] Output: Real-time power generation prediction of the wind-wave-current combined power generation device The addition of MLP reduces the prediction error of the model by 10% to 15% for scenarios with strong coupling of wind, waves and current (such as complex combinations of wind speed of 15m / s, tidal current speed of 3m / s, and wind direction of 90°).

[0112] S5: Model Closed-Loop Optimization (A Fusion Model of Four Algorithms: Data Update and Iterative Optimization) This step requires incorporating the MLP into closed-loop optimization, updating the dataset, and iteratively optimizing the parameters of the four models and the fusion weights. Input data includes the historical dataset from S1 and the four model predictions from S4. Actual real-time power The preprocessed features of S3 are used to output the updated dataset and the optimized fusion model.

[0113] Historical dataset update (add MLP-related records) The new sample must contain predictions from all four models, in the following format: ; (The original features have been corrected); (Actual power); to These are the real-time predictions from the four models (from S4).

[0114] The updated dataset is (k is the number of new samples).

[0115] Iterative optimization of the four-algorithm fusion model 1. Sub-model parameter adjustment: RF: Recalculate feature importance weights To adapt to dynamic changes in environmental parameters; XGBoost / GBRT: Adjusting the number of trees T and the learning rate For example, the period of high wind speed increases in summer. Up to 200; MLP: Optimizing the number of neurons in the hidden layer (e.g., adjusted from 16 to 32), learning rate and regularization coefficient This enhances the fitting of nonlinear coupling characteristics.

[0116] 2. Weight Update: Re-execute 5-fold cross-validation based on the updated data, calculate the RMSE of the four models, and update the weights. Ensure the weights of MLP in strongly coupled scenarios Dynamic improvement.

[0117] Optimize frequency and triggering conditions (adapt to four-model collaboration) Short-term optimization (per 1000 new samples): Only adjust the learning rate of XGBoost and GBRT and the hidden layer weights of MLP, while keeping the RF weights and fusion weights unchanged; Long-term optimization (per 10,000 new samples): Full retraining of all four models, including RF feature weights, MLP network structure and four-model fusion weights; Emergency optimization: When the MAPE of 10 consecutive predicted values ​​is greater than 7% (due to the addition of MLP, the threshold is relaxed from the original 5% but the actual accuracy is higher), full optimization is triggered immediately.

[0118] Output results: The updated historical dataset and the optimized RF-XGBoost-GBRT-MLP fusion model, through the dynamic collaboration of the four models, stabilize the long-term predicted MAPE within 4%, meeting the high-precision scheduling requirements of the wind-wave-current combined power generation system.

[0119] The RF-GBRT-based wind-wave-current combined power generation prediction system, applying any one of the above-mentioned RF-GBRT-based wind-wave-current combined power generation prediction methods, comprises: The data acquisition module is used to acquire historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical datasets, and to collect current environmental parameters in real time as real-time datasets. The model building module is used to construct the RF-XGBoost-GBRT fusion prediction model, including: training a random forest model to obtain normalized feature importance weights and random forest model prediction values ​​using historical environmental parameters as input and historical power generation data as output; weighting historical environmental parameters based on normalized feature importance weights to obtain weighted features; training an extreme gradient boosting model using weighted features as input and historical power generation data as output, outputting the initial prediction values ​​of extreme gradient boosting and calculating the residuals; training a gradient boosting regression tree model using historical environmental parameters as input and residuals as labels, outputting the corrected prediction values ​​of gradient boosting regression tree; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation; and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model. The real-time preprocessing module is used to correct outliers and normalize the real-time dataset. Then, based on the importance weights of the normalized features, it generates the weighted real-time features required for the extreme gradient boosting model and outputs the normalized real-time original features and weighted real-time features. The prediction module is used to input normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input weighted real-time features into the extreme gradient boosting model to obtain the real-time prediction values ​​of each model. Then, the prediction values ​​are weighted and summed according to the fusion weights to output the real-time power generation prediction value. The model optimization module is used to combine the normalized real-time original features, the real-time power generation prediction value and the corresponding actual real-time power generation into a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integrated prediction model.

Claims

1. A method for predicting combined wind, wave, and current power generation based on RF-GBRT, characterized in that, Includes the following steps: S1: Obtain historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical dataset, and collect current environmental parameters in real time as real-time dataset; S2: Construct an RF-XGBoost-GBRT fusion prediction model, including: using historical environmental parameters as input and historical power generation data as output, training a random forest model to obtain normalized feature importance weights and random forest model prediction values; weighting historical environmental parameters based on normalized feature importance weights to obtain extreme gradient boosting weighted features; using extreme gradient boosting weighted features as input and historical power generation data as output, training an extreme gradient boosting model, outputting preliminary extreme gradient boosting prediction values ​​and calculating residuals; using historical environmental parameters as input and residuals as labels, training a gradient boosting regression tree model, outputting gradient boosting regression tree corrected prediction values; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation, and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model; S3: Correct outliers and normalize the real-time dataset, then generate the weighted real-time features required for the extreme gradient boosting model based on the importance weights of the normalized features, and output the normalized real-time original features and weighted real-time features. S4: Input the normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input the weighted real-time features into the extreme gradient boosting model to obtain the real-time predicted values ​​of each model. Then, sum them according to the fusion weights to output the real-time power generation prediction value. S5: Combine the normalized real-time original features, the real-time power generation prediction value, and the corresponding actual real-time power generation to form a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integrated prediction model.

2. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S2, the training process of the random forest model includes: constructing multiple decision trees, obtaining training samples through bootstrap sampling, randomly selecting some historical environment parameters when splitting each tree, splitting based on the criterion of minimizing node impurity, calculating the total reduction in node impurity for each type of historical environment parameter as the feature importance weight, and then performing normalization processing.

3. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S2, the weighted features input to the extreme gradient boosting model are the product of historical environmental parameters and the importance weights of normalized features. The predicted values ​​are iteratively optimized through the gradient boosting strategy, so that the predicted values ​​gradually approach the difference between historical power generation data and the predicted values ​​of the random forest model.

4. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S2, the interaction process between the gradient boosting regression tree model and the extreme gradient boosting model is as follows: the training label of the gradient boosting regression tree model is the residual, and the unfitting bias in the residual is learned iteratively through the gradient boosting strategy. The output gradient boosting regression tree corrected prediction value is used to compensate for the prediction bias of the random forest model and the extreme gradient boosting model.

5. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S2, the process of determining the fusion weights is as follows: using 5-fold cross-validation, the root mean square error of the three models on the validation set is calculated. The fusion weights are inversely proportional to the root mean square error, and the sum of the fusion weights of the three models is 1.

6. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S1, the historical environmental parameters are wind speed, wind direction, and tidal current velocity; the real-time environmental parameters are real-time wind speed, real-time wind direction, and real-time tidal current velocity.

7. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S3, outlier correction uses the 3σ principle to determine abnormal real-time environmental parameters. Outliers are replaced by the mean of the parameters in the previous three normal times. In the 3σ principle, the mean and standard deviation are calculated based on the corresponding environmental parameters in the historical dataset.

8. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S5, the closed-loop optimization process includes short-term optimization and long-term optimization; short-term optimization involves adjusting the learning rate and number of trees of the extreme gradient boosting model and the gradient boosting regression tree model every 1000 new samples accumulated. In the long run, the optimization involves re-executing step 2 every 10,000 new samples to update the normalized feature importance weights, parameters of each sub-model, and fusion weights.

9. The method for predicting combined wind, wave, and current power generation based on RF-GBRT according to claim 1, characterized in that, In step S3, the normalization process includes: based on the minimum and maximum values ​​of each environmental parameter in the historical dataset, mapping the corrected real-time environmental parameters to the [0,1] interval to obtain normalized real-time features.

10. A wind-wave-current combined power generation prediction system based on RF-GBRT, characterized in that, The wind-wave-current combined power generation prediction method based on any one of claims 1 to 9, wherein the wind-wave-current combined power generation prediction system based on RF-GBRT comprises: The data acquisition module is used to acquire historical environmental parameters and corresponding historical power generation data of the wind-wave-current combined power generation device as historical datasets, and to collect current environmental parameters in real time as real-time datasets. The model building module is used to construct the RF-XGBoost-GBRT fusion prediction model, including: training a random forest model to obtain normalized feature importance weights and random forest model prediction values ​​using historical environmental parameters as input and historical power generation data as output; weighting historical environmental parameters based on normalized feature importance weights to obtain weighted features; training an extreme gradient boosting model using weighted features as input and historical power generation data as output, outputting the initial prediction values ​​of extreme gradient boosting and calculating the residuals; training a gradient boosting regression tree model using historical environmental parameters as input and residuals as labels, outputting the corrected prediction values ​​of gradient boosting regression tree; determining the fusion weights of the random forest model, extreme gradient boosting model, and gradient boosting regression tree model through cross-validation; and summing the prediction values ​​of the three models according to the fusion weights to obtain the fusion prediction model. The real-time preprocessing module is used to correct outliers and normalize the real-time dataset. Then, based on the importance weights of the normalized features, it generates the weighted real-time features required for the extreme gradient boosting model and outputs the normalized real-time original features and weighted real-time features. The prediction module is used to input normalized real-time raw features into the random forest model and the gradient boosting regression tree model, and input weighted real-time features into the extreme gradient boosting model to obtain the real-time prediction values ​​of each model. Then, the prediction values ​​are weighted and summed according to the fusion weights to output the real-time power generation prediction value. The model optimization module is used to combine the normalized real-time original features, the real-time power generation prediction value and the corresponding actual real-time power generation into a new sample, update the historical dataset, and repeat S2 to S4 based on the updated historical dataset to achieve closed-loop optimization of the fusion and integrated prediction model.

Images