Wind turbine power prediction method based on multi-modal feature fusion

Abstract

The invention provides a wind turbine power prediction method based on multi-modal feature fusion, and relates to the technical field of wind turbine power prediction.The method comprises the steps that multiple types of sensors are used for obtaining multi-source sensor data, and a standardized time sequence data set is generated; establishing a first observation matrix based on the standardized time sequence data set, and constructing a weight sensing fractional order adaptive genetic algorithm to optimize the first observation matrix to obtain a second observation matrix; performing feature fusion and noise reduction on the second observation matrix by applying extended Kalman filtering and combining the working state and physical modeling of the wind turbine; constructing a bidirectional long-short-term memory network, and performing power time sequence modeling based on the multi-dimensional noise reduction feature vector to obtain an output power predicted value; and feeding back an error between an output power prediction value and a true value to a weight sensing fractional order adaptive genetic algorithm to carry out parameter iterative optimization to obtain a third observation matrix, and obtaining a power prediction value of the wind turbine based on the third observation matrix.

Description

Technical Field

[0001] This invention relates to the field of wind turbine power prediction technology, and in particular to a wind turbine power prediction method based on multimodal feature fusion. Background Technology

[0002] As a core component of intelligent wind power system operation, wind turbine power prediction technology has undergone a systematic evolution in its research methods, from physical modeling to data-driven approaches. Traditional prediction methods mainly fall into two categories: physical models based on fluid dynamics principles and statistical models based on historical data analysis. Physical models construct numerical weather prediction systems by solving the Navier-Stokes equations, using atmospheric dynamics equations to numerically simulate the spatiotemporal distribution of wind speed, and then calculating the output power of wind turbines. These methods have clear physical mechanisms and high computational efficiency, but are limited by the inherent intermittent and random characteristics of wind speed, making it difficult to accurately capture the dynamic changes in wind power. Statistical models, on the other hand, mine the temporal correlation features in historical operational data, employing methods such as Autoregressive Integrated Moving Average (ARIMA) and Bayesian Model Average (BMA) to construct prediction models, enabling single-step or multi-step forward prediction. In recent years, with the rapid development of machine learning and deep learning technologies, data-driven methods based on artificial intelligence, with their powerful nonlinear mapping capabilities and adaptive learning characteristics, have shown significant advantages in the field of wind power prediction, gradually becoming a research hotspot and development trend in this area.

[0003] Zhang et al. proposed a method for short-term power prediction and uncertainty analysis of wind turbines based on Long Short-Term Memory (LSTM) networks and Gaussian Mixture Models (GMMs). This method first acquires meteorological data such as wind speed, wind direction, temperature, and air pressure through a Numerical Weather Prediction (NWP) system, using this data along with historical wind farm operation data as input. An LSTM network is then used to construct a multi-time-step power prediction model. The LSTM processes the time-series data through input gates, forget gates, and output gates to capture the long-term dependencies of the power sequence. During prediction, the processed feature vectors are input into the trained LSTM network, which outputs power prediction values ​​for multiple future time steps. This method uses a Gaussian Mixture Model to probabilistically model the prediction error distribution and estimates the mixture Gaussian distribution parameters using the Expectation-Maximization (EM) algorithm, achieving quantitative analysis of prediction uncertainty. While this method effectively captures the long-term dependencies of the power sequence using the LSTM network and achieves uncertainty quantification through the Gaussian Mixture Model, it lacks an effective feature fusion mechanism when handling high-dimensional features, making it difficult to fully utilize the complementary information from multimodal sensor data. Meanwhile, this method does not consider the impact of sensor noise on prediction accuracy, and in practical engineering applications, prediction performance is prone to decline due to data quality issues.

[0004] The paper "Wind Turbine Power Output Prediction Using a New Hybrid Neuro-Evolutionary Method" proposes a hybrid neuro-evolutionary method for predicting wind turbine power output. This method comprises three stages: a preprocessing stage employs K-means clustering and an autoencoder to detect and filter noise in SCADA measurement data; a decomposition stage proposes a hybrid variational mode decomposition (HVMD) method, combining a greedy Nelder-Mead search algorithm and an adaptive stochastic local search to fine-tune the VMD hyperparameters; and a prediction stage uses an adaptive differential evolution algorithm and a sine / cosine optimization method as hyperparameter optimizers, combined with a long short-term memory network, to establish a wind turbine power curve model. This framework achieves data flow between modules through standardized interfaces, uses permutation entropy to evaluate the complexity of the decomposed sub-signals, and achieves 10-minute and 1-hour advance predictions for onshore wind farms in Sweden. Although this method introduces data preprocessing and variational mode decomposition techniques, the evolutionary algorithm used is inefficient in high-dimensional parameter space searches and lacks domain knowledge guidance specific to wind power characteristics, resulting in slow convergence and difficulty in meeting real-time prediction requirements.

[0005] While existing wind turbine power prediction technologies have made some progress in both theory and application, several key performance defects remain, limiting their prediction accuracy and engineering practicality under complex operating conditions. A thorough analysis of the common problems in existing technologies reveals three main aspects: First, existing prediction systems have significant shortcomings in multimodal signal fusion. The strong nonlinear coupling characteristics of thermodynamics, mechanical dynamics, and electrical parameters make it difficult for traditional methods to accurately model these parameters, and robust solutions for the effective fusion of multi-source heterogeneous data under sensor noise interference are still lacking. Second, existing optimization algorithms generally suffer from parameter sensitivity issues. The setting of key parameters such as crossover probability and mutation probability is highly dependent on empirical tuning; inappropriate parameter configuration can easily lead to the algorithm getting stuck in local optima or having excessively slow convergence. Finally, the balance between the real-time constraints and accuracy requirements of prediction models in industrial scenarios has not been systematically optimized. Existing algorithms often suffer from high computational complexity, making them difficult to meet the needs of edge deployment and limiting their widespread application in actual wind farms. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a wind turbine power prediction method based on multimodal feature fusion, comprising the following steps:

[0007] For wind turbines, multiple types of sensors are used to acquire multi-source sensor data. The acquired multi-source sensor data is preprocessed to generate a standardized time-series dataset.

[0008] A first observation matrix is ​​established based on a standardized time-series dataset. A weight-aware fractional adaptive genetic algorithm WAFA-GA is constructed to optimize the first observation matrix, resulting in a second observation matrix.

[0009] By applying extended Kalman filtering and combining the working state and physical modeling of the wind turbine, feature fusion and noise reduction are performed on the second observation matrix to obtain a multi-dimensional noise-reduced feature vector.

[0010] A bidirectional long short-term memory network Bi-LSTM is constructed, and wind turbine power time series modeling is performed based on multi-dimensional noise-reduced feature vectors to obtain the output power prediction value.

[0011] The error between the predicted output power value and the actual value is fed back to the weighted fractional adaptive genetic algorithm WAFA-GA for parameter iterative optimization, resulting in the third observation matrix. The predicted power value of the wind turbine is then obtained based on the third observation matrix.

[0012] Furthermore, the sensors include various categories such as gearbox temperature sensors, bearing temperature sensors, tower vibration acceleration sensors, blade pitch angular position sensors, electrical parameter sensors, and wind condition monitoring sensors; multi-source sensor data includes thermodynamic parameters, mechanical dynamic parameters, and electrical parameters.

[0013] The collected multi-source sensor data is preprocessed, including: removing null records caused by sensor failures, using median filtering to eliminate impulse noise, solving the multi-source signal transmission delay problem through a timestamp alignment mechanism, and generating a standardized time-series dataset.

[0014] Furthermore, a first observation matrix is ​​established based on the standardized time-series dataset, specifically as follows:

[0015] Using a bidirectional long short-term memory network (Bi-LSTM) as the feature contribution evaluation model, the contribution weights of each type of sensor are calculated, and the contribution weights of each type of sensor are concatenated to obtain the first observation matrix.

[0016] Furthermore, the constructed weight-aware fractional-order adaptive genetic algorithm WAFA-GA includes selection, crossover, and mutation operations, and establishes a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, a multi-objective fitness function, and a hierarchy of solutions based on non-dominated sorting.

[0017] Establish a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, including:

[0018] A fractional-order sigmoid regulator based on the Mittag-Leffler function is introduced, using the Mittag-Leffler function as the parameter adjustment function, as shown in the following formula:

[0019]

[0020] in, For the Gamma function, The attenuation coefficient is... For fractional-order parameters;

[0021] Based on the parameter adjustment function, a fractional-order adaptive adjustment rule for adjusting the crossover probability and mutation probability is established, as shown in the following formula:

[0022]

[0023]

[0024] in, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first Replace the optimal fitness value, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first The optimal fitness value;

[0025] Select multiple optimization objectives and set at least one constraint objective. Establish a multi-objective fitness function based on the multiple optimization objectives and at least one constraint objective.

[0026] Establish a hierarchical system for solutions based on non-dominated sorting to stratify and rank individuals in the population, including:

[0027] Based on Pareto dominance, individuals in the population are hierarchically ranked: For two candidate individuals A and B in the observation matrix, individual A is said to dominate individual B if it is no worse than individual B on all optimization objectives and better than individual B on at least one optimization objective; all individuals not dominated by any other individual are identified as the first frontier. Then, from the remaining individuals, non-dominated individuals are identified as the second frontier. ; and so on, iterating through the selection process until all individuals have been sorted hierarchically.

[0028] Furthermore, the first observation matrix is ​​optimized using the following method:

[0029] The population size and maximum number of generations are set. The first observation matrix is ​​optimized using the weight-aware fractional adaptive genetic algorithm WAFA-GA. Selection, crossover, and mutation operations are performed. The crossover and mutation probabilities are dynamically adjusted according to the multi-objective fitness function and the fractional adaptive parameter adjustment mechanism. The population individuals are hierarchically sorted by establishing a solution hierarchy through non-dominated sorting. When the weight-aware fractional adaptive genetic algorithm converges or reaches the maximum number of generations, the second observation matrix is ​​obtained.

[0030] Furthermore, extended Kalman filtering is applied, combined with the operating state and physical modeling of the wind turbine, to perform feature fusion and noise reduction on the second observation matrix, including:

[0031] The wind turbine system is modeled as a nonlinear dynamic system. The second observation matrix is ​​converted into a multidimensional system state vector. The control matrix is ​​defined, and a state-space model of the wind turbine system is constructed, including state evolution equations and observation equations.

[0032] The extended Kalman filter is used to recursively estimate the state of the wind turbine system. The state of the wind turbine system is estimated in two stages: prediction and update, and the updated state estimate is obtained.

[0033] Feature extraction is performed on the updated state estimate to generate a multidimensional denoised feature vector.

[0034] Furthermore, the extended Kalman filter is used to recursively estimate the wind turbine system state. The wind turbine system state is estimated through a two-stage process of prediction and update. The specific method is as follows:

[0035] In the prediction phase, based on the wind turbine system state estimate and control input vector at the previous moment, the wind turbine system state estimate and covariance at the current moment are predicted.

[0036] During the update phase, the Kalman gain is calculated to correct the predicted state estimate and covariance of the wind turbine system at the current moment, resulting in the updated state estimate.

[0037] Furthermore, feature extraction is performed on the updated state estimates, specifically as follows:

[0038] Based on the power characteristic curve constraint of the wind turbine, the characteristics of the thermodynamic parameters contained in the updated state estimate are normalized by temperature threshold.

[0039] Based on the principles of tower structure dynamics, vibration frequency domain filtering is applied to the characteristics of the mechanical dynamic parameters contained in the updated state estimate to eliminate abnormal components near the resonance frequency.

[0040] Based on the power balance constraints of the electrical system, the characteristics of the electrical parameters included in the updated state estimate are normalized to maintain the relative proportions of voltage and current parameters.

[0041] Furthermore, a bidirectional long short-term memory network (Bi-LSTM) is constructed, and power time-series modeling is performed based on multi-dimensional denoising feature vectors to obtain the output power prediction value. The specific method is as follows:

[0042] A two-layer Bi-LSTM network is constructed. Each Bi-LSTM network contains two sub-networks: a forward LSTM and a backward LSTM. Each sub-network contains several hidden units. The first layer Bi-LSTM network performs preliminary feature extraction on the multi-dimensional noise reduction feature vector. The second layer Bi-LSTM performs deep feature fusion on the feature extraction results output by the first layer Bi-LSTM network.

[0043] A bidirectional long short-term memory (Bi-LSTM) network is used to extract features from the multidimensional denoised feature vector. The multidimensional denoised feature vector is taken as input and processed by the forward LSTM and backward LSTM of the first layer Bi-LSTM network to process the forward and backward information of the multidimensional denoised feature vector, respectively. The forward LSTM captures the influence of the historical state on the current state, and the backward LSTM captures the influence of the future state on the current state. The outputs of the forward LSTM and backward LSTM are concatenated at each time step and then input into the second layer Bi-LSTM network to obtain the fused features.

[0044] The fused features output by the two-layer Bi-LSTM network are mapped to power prediction values ​​through a fully connected layer.

[0045] Furthermore, the error between the predicted output power value and the actual value is fed back to the weight-aware fractional adaptive genetic algorithm WAFA-GA for parameter iterative optimization. The specific method is as follows:

[0046] Calculate the prediction error information between the output power prediction value and the actual value, including the mean absolute error (MAE) and root mean square error (RMSE) of the power prediction value, and determine whether it meets the preset accuracy requirements; if it does not meet the requirements and does not exceed the maximum number of iterations, perform parameter feedback adjustment; otherwise, output the final prediction result.

[0047] Parameter feedback adjustments include:

[0048] The prediction error information is fed back to the weight-aware fractional adaptive genetic algorithm WAFA-GA for parameter optimization, resulting in the third observation matrix, which includes:

[0049] Based on the prediction error information, adjust the range of values ​​for the fractional-order parameter. When the prediction error is greater than the set threshold, increase the fractional-order parameter to speed up the parameter adjustment. When the prediction error is less than the set threshold, decrease the value to improve convergence stability.

[0050] Based on the time-series distribution characteristics of the error, the decay coefficient is dynamically adjusted to optimize the decay characteristics of the Mittag-Leffler function, thereby achieving adaptive adjustment of crossover probability and mutation probability.

[0051] Based on the prediction accuracy, the weight coefficients in the multi-objective fitness function are recalculated to balance the importance weights of multiple optimization objectives in the first observation matrix optimization process.

[0052] The predicted power of the wind turbine is obtained based on the third observation matrix.

[0053] The beneficial effects of adopting the above technical solution are as follows: The wind turbine power prediction method based on multimodal feature fusion provided by this invention constructs a weighted fractional-order adaptive genetic algorithm WAFA-GA to achieve heuristic population initialization based on feature contribution evaluation, and introduces a Mittag-Leffler function to dynamically adjust the genetic parameters, effectively overcoming the inherent defects of traditional optimization algorithms such as high parameter sensitivity and easy getting trapped in local optima. Simultaneously, this invention designs an extended Kalman filter-driven multimodal feature fusion unit, reconstructing the denoised feature vector through physical constraints, solving the robustness problem of heterogeneous data fusion under sensor noise interference. Furthermore, this invention uses a bidirectional long short-term memory network to capture forward and backward dynamic dependencies, overcoming the bottleneck of insufficient modeling ability for nonlinear temporal correlations in traditional models. Through a closed-loop feedback mechanism of "optimization-fusion-prediction" to achieve deep collaboration between modules, this invention significantly improves the prediction accuracy of wind turbine output power while meeting the real-time constraints of industrial scenarios, providing a technical solution for wind power system condition monitoring that combines theoretical innovation and engineering practical value. Attached Figure Description

[0054] Figure 1 A flowchart of a wind turbine power prediction method based on multimodal feature fusion provided in an embodiment of the present invention;

[0055] Figure 2 The flowchart of the weight-aware fractional-order adaptive genetic algorithm WAFA-GA provided in the embodiments of the present invention. Detailed Implementation

[0056] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0057] This embodiment presents a wind turbine power prediction method based on multimodal feature fusion, such as... Figure 1 As shown, it includes the following steps:

[0058] Step 1: Use various types of sensors to acquire multi-source sensor data for wind turbines, preprocess the acquired multi-source sensor data, and generate a standardized time-series dataset;

[0059] Various types of sensors include gearbox temperature sensors, bearing temperature sensors, tower vibration acceleration sensors, blade pitch angular position sensors, electrical parameter sensors, and wind condition monitoring sensors; multi-source sensor data includes thermodynamic parameters, mechanical dynamic parameters, and electrical parameters;

[0060] The collected multi-source sensor data is preprocessed, including: removing null records caused by sensor failures, eliminating impulse noise using median filtering, solving the multi-source signal transmission delay problem through a timestamp alignment mechanism, and generating a standardized time-series dataset.

[0061] In this embodiment, multi-source sensor data, including thermodynamic parameters, mechanical dynamic parameters, and electrical parameters, are acquired in real time through the wind farm SCADA system. The thermodynamic parameters include gearbox bearing temperature and coolant temperature; the mechanical dynamic parameters include tower three-dimensional vibration acceleration and pitch angle dynamic parameters; and the electrical parameters include generator winding current and DC bus voltage, resulting in 71-dimensional heterogeneous sensor data. The data acquisition frequency in this embodiment is 10 minutes / time, and the time span is from January 1, 2019 to August 14, 2021, resulting in 136,728 valid records.

[0062] Step 2: Establish the first observation matrix based on the standardized time series dataset, construct the weight-aware fractional adaptive genetic algorithm WAFA-GA, optimize the first observation matrix, and obtain the second observation matrix;

[0063] The observation matrix is ​​constructed based on a standardized time-series dataset. The specific method is as follows:

[0064] A bidirectional long short-term memory network (Bi-LSTM) is used as the feature contribution evaluation model to calculate the contribution weights of each sensor category. The contribution weights of each sensor category are calculated using the following formula:

[0065]

[0066]

[0067] in, for Hidden state of Bi-LSTM network at different times This is the hidden layer bias vector. For trainable attention weight matrix, For time step, Indicates the first Sensors of different categories Attention score at any moment For the first Contribution weights for each sensor category;

[0068] The contribution weights of each type of sensor are concatenated to obtain the first observation matrix. As shown in the formula below:

[0069]

[0070] The population is heuristically initialized using a feature contribution evaluation model. The first observation matrix obtained through heuristic initialization provides a better starting point compared to random initialization.

[0071] Step 2.2: Construct the weight-aware fractional adaptive genetic algorithm WAFA-GA, as follows: Figure 2 As shown, the weight-aware fractional-order adaptive genetic algorithm WAFA-GA includes selection, crossover, and mutation operations, and establishes a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, a multi-objective fitness function, and a hierarchy of solutions based on non-dominated sorting.

[0072] Establish a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, including:

[0073] A fractional-order sigmoid regulator based on the Mittag-Leffler function is introduced, using the Mittag-Leffler function as the parameter adjustment function, as shown in the following formula:

[0074]

[0075] in, For the Gamma function, The attenuation coefficient is... For fractional-order parameters;

[0076] Based on the parameter adjustment function, a fractional-order adaptive adjustment rule for adjusting the crossover probability and mutation probability is established, as shown in the following formula:

[0077]

[0078]

[0079] in, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first Replace the optimal fitness value, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first The optimal fitness value;

[0080] The specific method for establishing a multi-objective fitness function is as follows:

[0081] Selecting the mean absolute error Root mean square error As the optimization objective, mean absolute error and root mean square error As shown in the formula below:

[0082]

[0083]

[0084] in, For the first The actual output power of each sample For the first The predicted power of each sample The number of samples;

[0085] Set the constraint target to floating-point operation constraint, as shown in the following formula:

[0086]

[0087] A multi-objective fitness function is established based on the optimization objective and floating-point operation constraints. As shown in the formula below:

[0088]

[0089] in, As a weighting coefficient, in this embodiment, it is set to... .

[0090] Establish a hierarchical system for solutions based on non-dominated sorting to stratify and rank individuals in the population, including:

[0091] Based on Pareto dominance, individuals in the population are stratified and ranked: For two candidate individuals A and B in the observation matrix, if individual A is not inferior to individual B on both the mean absolute error f1 and the root mean square error f2, and is superior to individual B on at least one optimization objective, then individual A is said to dominate individual B; all individuals not dominated by any other individual are identified as the first frontier. Then, from the remaining individuals, non-dominated individuals are identified as the second frontier. ; and so on, iterating through the selection process until all individuals have been sorted hierarchically;

[0092] Step 2.3: Set the population size and maximum number of generations, and use the weight-aware fractional adaptive genetic algorithm WAFA-GA to optimize the first observation matrix to obtain the second observation matrix;

[0093] In this embodiment, the population size is set to 50, the maximum number of generations is 70, and the weight coefficient is 0.3. Selection, crossover, and mutation operations are performed. The crossover and mutation probabilities are dynamically adjusted according to the multi-objective fitness function and the fractional-order adaptive parameter adjustment mechanism. A hierarchical system of solutions is established through non-dominated sorting to stratify and sort the individuals in the population. When the weight-aware fractional-order adaptive genetic algorithm converges or reaches the maximum number of generations, the second observation matrix is ​​output. ;

[0094] In this embodiment, a hierarchy of solutions is established based on non-dominated sorting. An elite retention strategy coordinates the dual-objective optimization of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), ensuring that the floating-point operation constraint FLOPs ≤ 10 is met. 8 Under the premise of maintaining population diversity to suppress premature convergence, and driving the parameter optimization process to converge efficiently, an observation matrix that balances prediction accuracy and computational efficiency is ultimately generated. .

[0095] Step 3: Apply extended Kalman filtering, combined with the working state and physical modeling of the wind turbine, to perform feature fusion and noise reduction on the second observation matrix to obtain a multi-dimensional noise-reduced feature vector;

[0096] The wind turbine system is modeled as a nonlinear dynamic system, and the second observation matrix is ​​used. Convert to 3D system state vector Define the control matrix The pitch angle adjustment and power constraint of the wind turbine system are mapped to the state space to construct a state space model of the wind turbine system, including state evolution equations and observation equations, as shown in the following formulas:

[0097]

[0098]

[0099] in, To control the input vector, To control the dimension of the input vector, These are the three pitch angle adjustment values ​​for the wind turbine. Due to internal power limitations, Due to external power limitations, For the output power of the wind turbine system, For process noise, To observe the noise, For nonlinear observation functions, It is a nonlinear state transition function. The process noise covariance matrix is... To observe the noise covariance matrix;

[0100] Process noise covariance matrix This is used to describe the uncertainties in the state transition process of a wind turbine system, including the randomness of gearbox bearing temperature changes, environmental disturbances caused by tower vibration, and measurement errors of electrical parameters; the observation noise covariance matrix. Used to characterize the uncertainty of power output measurements;

[0101] The extended Kalman filter is used to recursively estimate the state of the wind turbine system. The estimation of the wind turbine system state is carried out in two stages: prediction and update, to obtain the updated state estimate. The specific method is as follows:

[0102] During the prediction phase, the wind turbine system state estimate is used as a basis. and control input vector Predict the current state estimate of the wind turbine system. Covariance As shown in the formula below:

[0103]

[0104]

[0105] in, The state transition Jacobian matrix;

[0106] During the update phase, the Kalman gain is calculated. Correct the predicted current state estimate of the wind turbine system. Covariance The updated state estimate is obtained. As shown in the formula below:

[0107]

[0108]

[0109]

[0110] in, To observe the Jacobian matrix, For Kalman gain, It is the identity matrix. The updated state estimate, This is the updated covariance matrix;

[0111] Updated state estimates Feature extraction is performed, and the updated state estimate is reconstructed using a physical constraint mechanism. The thermodynamic, mechanical, and electrical features contained therein are mapped to a unified feature space to generate... The specific method for denoising feature vectors is as follows:

[0112] Based on the power characteristic curve constraints of the wind turbine, the updated state estimate is... The thermodynamic parameters contained therein are subjected to temperature threshold normalization to ensure that parameters such as gearbox temperature and bearing temperature are within a reasonable operating range.

[0113] Based on the principles of tower structure dynamics, the updated state estimate is... The characteristics of the mechanical dynamic parameters contained therein are subjected to vibration frequency domain filtering to eliminate abnormal components near the resonance frequency;

[0114] Based on the power balance constraints of the electrical system, the updated state estimate is... The electrical parameters contained therein are normalized in amplitude to maintain the relative proportional relationship of voltage and current parameters, thereby achieving physical consistency reconstruction of multimodal features.

[0115] Step 4: Construct a bidirectional long short-term memory network Bi-LSTM, and perform time series modeling of wind turbine power based on multi-dimensional noise-reduced feature vectors to obtain the predicted output power value;

[0116] Construct a two-layer Bi-LSTM network, each layer of which contains two sub-networks: a forward LSTM and a backward LSTM. Each sub-network contains 128 hidden units. The first layer of the Bi-LSTM network... The first layer of Bi-LSTM performs preliminary feature extraction on the denoised feature vector, and the second layer of Bi-LSTM performs deep feature fusion on the feature extraction results output by the first layer of Bi-LSTM network.

[0117] Using a bidirectional long short-term memory network Bi-LSTM Feature extraction is performed on the 3D noise reduction feature vector, and... The denoised feature vector is taken as input and processed by the forward LSTM and backward LSTM of the first layer Bi-LSTM network, respectively. The forward and backward information of the denoising feature vector are obtained by using a forward LSTM to capture the influence of the historical state on the current state and a backward LSTM to capture the influence of the future state on the current state. The outputs of the forward and backward LSTMs are concatenated at each time step and then input into the second layer Bi-LSTM network to achieve deep bidirectional state fusion and obtain fused features.

[0118] The fused features output by the two-layer Bi-LSTM network are mapped to power prediction values ​​through a fully connected layer;

[0119] In this embodiment, the Adam optimizer is used to minimize the prediction error, the learning rate is set to 0.001, the mean squared error is used as the loss function during training, and the network parameters are updated through the backpropagation algorithm.

[0120] Step 5: Feed back the error between the predicted output power value and the actual value to the weight-aware fractional adaptive genetic algorithm WAFA-GA for parameter iterative optimization to obtain the third observation matrix, and obtain the predicted power value of the wind turbine based on the third observation matrix.

[0121] Calculate the prediction error information between the output power prediction value and the actual value, including the mean absolute error (MAE) and root mean square error (RMSE) of the power prediction value, and determine whether it meets the preset accuracy requirements; if it does not meet the requirements and does not exceed the maximum number of iterations, perform parameter feedback adjustment; otherwise, output the final prediction result.

[0122] Parameter feedback adjustments include:

[0123] The prediction error information is fed back to the weight-aware fractional adaptive genetic algorithm WAFA-GA for parameter optimization, resulting in the third observation matrix. ,include:

[0124] Adjust the fractional-order parameters based on the prediction error information. The value range is increased when the prediction error exceeds the set threshold. The value is reduced to speed up parameter adjustment; when the prediction error is less than the set threshold, the value is decreased. Values ​​are used to improve convergence stability;

[0125] Based on the error time series distribution characteristics, the attenuation coefficient is dynamically adjusted. Optimize the decay characteristics of the Mittag-Leffler function to achieve crossover probability and mutation probability Adaptive adjustment;

[0126] Based on the prediction accuracy, the weight coefficients in the multi-objective fitness function are recalculated. Balanced mean absolute error and root mean square error Importance weights in the optimization process of the first observation matrix;

[0127] In this embodiment, the parameters of the weight-aware fractional-order adaptive genetic algorithm WAFA-GA are optimized to form a closed-loop iterative mechanism of "optimization-fusion-prediction". The predicted power value of the wind turbine is obtained based on the third observation matrix. Based on the 71-dimensional heterogeneous sensor data constructed in this embodiment, the wind turbine output power prediction method based on multimodal feature fusion provided in this embodiment is used to predict the wind turbine output power. The specific detection indicators and experiments are shown in Table 1.

[0128] Table 1. Detection Indicators of Wind Turbine Power Prediction Method Based on Multimodal Feature Fusion

[0129]

[0130] The weight-aware fractional adaptive genetic algorithm WAFA-GA was used to optimize key performance indicators. The specific test indicators and experimental results are shown in Table 2.

[0131] Table 2. Detection metrics of the weight-aware fractional adaptive genetic algorithm WAFA-GA

[0132]

[0133] This embodiment utilizes a wind turbine power prediction method based on multimodal feature fusion to predict the output power of operating wind turbines, significantly improving prediction accuracy and effectively predicting wind turbine output power. This invention achieves a revolutionary breakthrough in wind turbine power prediction accuracy through the synergistic effect of multimodal feature fusion and intelligent optimization. Experimental data shows that, under standard test conditions, the mean absolute error (MAE) decreased from 2.87006 kW in the traditional physical model to 0.2553 kW, a reduction of 90.44%; the root mean square error (RMSE) decreased from 3.39902 kW to 0.3485 kW, a reduction of 89.10%. Simultaneously, the coefficient of determination approaches the theoretical limit, reaching 0.99999999, indicating a very high correlation between the predicted and actual values. This accuracy improvement stems from the accurate modeling of the multi-physics coupling relationship in this invention, particularly the intelligent fusion of sensor data achieved through an extended Kalman filter, which effectively eliminates the influence of measurement noise and system disturbances.

[0134] The extended Kalman filter-driven feature fusion unit designed in this embodiment can simultaneously process 71-dimensional sensor data including thermodynamic, mechanical, and electrical parameters. It achieves unified representation and noise suppression of heterogeneous data through an adaptive state estimation framework. The innovation of this fusion mechanism lies in preserving the dimensional differences of each physical quantity, truly reflecting the distribution characteristics of the multi-dimensional state space. Furthermore, through the design of a dynamically observed Jacobian matrix H and a control matrix B, it realizes the mapping of different types of sensor data, effectively solving the sensor noise interference problem by utilizing the multi-source heterogeneous data fusion capability. Experimental results show that after feature fusion processing, the system can effectively identify and suppress irregular outliers caused by sensor faults and transmission anomalies, significantly improving the robustness of the prediction model.

[0135] The WAFA-GA algorithm proposed in this embodiment achieves a breakthrough in parameter optimization efficiency. The algorithm boasts fast convergence speed and strong global search capability. Through a heuristic population initialization strategy based on feature contribution, the algorithm can quickly approach a high-quality solution space from the early stages of the search, avoiding a large amount of ineffective searching caused by traditional random initialization. The fractional-order adaptive adjustment mechanism enables the algorithm to dynamically balance global exploration and local exploitation, converging to the optimal solution within 28 generations of evolution. Compared to classic methods such as index-based evolutionary algorithms and multi-objective evolutionary decomposition algorithms, the convergence speed is improved by 44% to 50%. More importantly, the algorithm's premature convergence rate is only 14.29%, far lower than the 36% to 74% of the comparative algorithms, indicating its excellent global search capability and ability to escape local optima. The peak population standard deviation reaches 0.0122, an improvement of 64.7% to 79.8% compared to traditional algorithms, fully demonstrating the algorithm's advantage in maintaining population diversity. This efficient optimization performance ensures that the system can find the optimal parameter configuration with limited computational resources.

[0136] This embodiment fully considers the hardware limitations of wind farm edge computing devices. Through careful architecture design and algorithm optimization, the overall floating-point computation is strictly controlled to within 10^8. This constraint ensures that the prediction model can run stably on resource-constrained embedded systems and edge servers without relying on expensive high-performance computing clusters. Simultaneously, by employing compact variable naming and efficient data structure design, memory consumption and data transfer overhead are minimized, meeting industrial real-time requirements with low computational resource consumption. This provides a feasible technical path for the large-scale engineering application of wind power prediction systems.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for predicting wind turbine power based on multimodal feature fusion, characterized in that, Includes the following steps: For wind turbines, multiple types of sensors are used to acquire multi-source sensor data. The acquired multi-source sensor data is preprocessed to generate a standardized time-series dataset. A first observation matrix is ​​established based on a standardized time-series dataset. A weight-aware fractional adaptive genetic algorithm WAFA-GA is constructed to optimize the first observation matrix, resulting in a second observation matrix. By applying extended Kalman filtering and combining the working state and physical modeling of the wind turbine, feature fusion and noise reduction are performed on the second observation matrix to obtain a multi-dimensional noise-reduced feature vector. A bidirectional long short-term memory network Bi-LSTM is constructed, and wind turbine power time series modeling is performed based on multi-dimensional noise-reduced feature vectors to obtain the output power prediction value. The error between the predicted output power value and the actual value is fed back to the weighted fractional adaptive genetic algorithm WAFA-GA for parameter iterative optimization, resulting in the third observation matrix. The predicted power value of the wind turbine is then obtained based on the third observation matrix.

2. The wind turbine power prediction method based on multimodal feature fusion according to claim 1, characterized in that, Various types of sensors include gearbox temperature sensors, bearing temperature sensors, tower vibration acceleration sensors, blade pitch angular position sensors, electrical parameter sensors, and wind condition monitoring sensors; multi-source sensor data includes thermodynamic parameters, mechanical dynamic parameters, and electrical parameters; The collected multi-source sensor data is preprocessed, including: removing null records caused by sensor failures, using median filtering to eliminate impulse noise, solving the multi-source signal transmission delay problem through a timestamp alignment mechanism, and generating a standardized time-series dataset.

3. The wind turbine power prediction method based on multimodal feature fusion according to claim 2, characterized in that, The first observation matrix is ​​established based on a standardized time-series dataset. The specific method is as follows: Using a bidirectional long short-term memory network (Bi-LSTM) as the feature contribution evaluation model, the contribution weights of each type of sensor are calculated, and the contribution weights of each type of sensor are concatenated to obtain the first observation matrix.

4. The wind turbine power prediction method based on multimodal feature fusion according to claim 3, characterized in that, The constructed weight-aware fractional-order adaptive genetic algorithm WAFA-GA includes selection, crossover, and mutation operations. It establishes a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, a multi-objective fitness function, and a hierarchy of solutions based on non-dominated sorting. Establish a fractional-order adaptive parameter adjustment mechanism for adjusting crossover and mutation probabilities, including: A fractional-order sigmoid regulator based on the Mittag-Leffler function is introduced, using the Mittag-Leffler function as the parameter adjustment function, as shown in the following formula: ； in, For the Gamma function, The attenuation coefficient is... For fractional-order parameters; Based on the parameter adjustment function, a fractional-order adaptive adjustment rule for adjusting the crossover probability and mutation probability is established, as shown in the following formula: ； ； in, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first Replace the optimal fitness value, and The first Crossover probability and mutation probability of each generation For the first Standard deviation of population fitness For the first The optimal fitness value; Select multiple optimization objectives and set at least one constraint objective. Establish a multi-objective fitness function based on the multiple optimization objectives and at least one constraint objective. Establish a hierarchical system for solutions based on non-dominated sorting to stratify and rank individuals in the population, including: Based on Pareto dominance, individuals in the population are hierarchically ranked: For two candidate individuals A and B in the observation matrix, individual A is said to dominate individual B if it is no worse than individual B on all optimization objectives and better than individual B on at least one optimization objective; all individuals not dominated by any other individual are identified as the first frontier. Then, from the remaining individuals, non-dominated individuals are identified as the second frontier. ; and so on, iterating through the selection process until all individuals have been sorted hierarchically.

5. The wind turbine power prediction method based on multimodal feature fusion according to claim 4, characterized in that, The first observation matrix is ​​optimized using the following method: The population size and maximum number of generations are set. The first observation matrix is ​​optimized using the weight-aware fractional adaptive genetic algorithm WAFA-GA. Selection, crossover, and mutation operations are performed. The crossover and mutation probabilities are dynamically adjusted according to the multi-objective fitness function and the fractional adaptive parameter adjustment mechanism. The population individuals are hierarchically sorted by establishing a solution hierarchy through non-dominated sorting. When the weight-aware fractional adaptive genetic algorithm converges or reaches the maximum number of generations, the second observation matrix is ​​obtained.

6. The wind turbine power prediction method based on multimodal feature fusion according to claim 5, characterized in that, Extended Kalman filtering is applied, combined with the operating state and physical modeling of the wind turbine, to perform feature fusion and noise reduction on the second observation matrix, including: The wind turbine system is modeled as a nonlinear dynamic system. The second observation matrix is ​​converted into a multidimensional system state vector. The control matrix is ​​defined, and a state-space model of the wind turbine system is constructed, including state evolution equations and observation equations. The extended Kalman filter is used to recursively estimate the state of the wind turbine system. The state of the wind turbine system is estimated in two stages: prediction and update, and the updated state estimate is obtained. Feature extraction is performed on the updated state estimate to generate a multidimensional denoised feature vector.

7. The wind turbine power prediction method based on multimodal feature fusion according to claim 6, characterized in that, The extended Kalman filter is used to recursively estimate the wind turbine system state. The estimation of the wind turbine system state is carried out in two stages: prediction and update. The specific method is as follows: In the prediction phase, based on the wind turbine system state estimate and control input vector at the previous moment, the wind turbine system state estimate and covariance at the current moment are predicted. During the update phase, the Kalman gain is calculated to correct the predicted state estimate and covariance of the wind turbine system at the current moment, resulting in the updated state estimate.

8. The wind turbine power prediction method based on multimodal feature fusion according to claim 7, characterized in that, Feature extraction is performed on the updated state estimates, specifically using the following method: Based on the power characteristic curve constraint of the wind turbine, the characteristics of the thermodynamic parameters contained in the updated state estimate are normalized by temperature threshold. Based on the principles of tower structure dynamics, vibration frequency domain filtering is applied to the characteristics of the mechanical dynamic parameters contained in the updated state estimate to eliminate abnormal components near the resonance frequency. Based on the power balance constraints of the electrical system, the characteristics of the electrical parameters included in the updated state estimate are normalized to maintain the relative proportions of voltage and current parameters.

9. The wind turbine power prediction method based on multimodal feature fusion according to claim 8, characterized in that, A bidirectional long short-term memory (Bi-LSTM) network is constructed, and power time-series modeling is performed based on multi-dimensional noise-reduced feature vectors to obtain the predicted output power value. The specific method is as follows: Construct a two-layer Bi-LSTM network. Each layer of the Bi-LSTM network contains two sub-networks: a forward LSTM and a backward LSTM. Each sub-network contains several hidden units. The first layer of Bi-LSTM network performs preliminary feature extraction on the multidimensional noise reduction feature vector, and the second layer of Bi-LSTM performs deep feature fusion on the feature extraction results output by the first layer of Bi-LSTM network. A bidirectional long short-term memory (Bi-LSTM) network is used to extract features from the multidimensional denoised feature vector. The multidimensional denoised feature vector is taken as input and processed by the forward LSTM and backward LSTM of the first layer Bi-LSTM network to process the forward and backward information of the multidimensional denoised feature vector, respectively. The forward LSTM captures the influence of the historical state on the current state, and the backward LSTM captures the influence of the future state on the current state. The outputs of the forward LSTM and backward LSTM are concatenated at each time step and then input into the second layer Bi-LSTM network to obtain the fused features. The fused features output by the two-layer Bi-LSTM network are mapped to power prediction values ​​through a fully connected layer.

10. The wind turbine power prediction method based on multimodal feature fusion according to claim 9, characterized in that, The error between the predicted and actual output power values ​​is fed back to the weight-aware fractional adaptive genetic algorithm WAFA-GA for iterative parameter optimization. The specific method is as follows: Calculate the prediction error information between the predicted output power value and the actual value, including the mean absolute error (MAE) and root mean square error (RMSE) of the predicted power value, and determine whether the preset accuracy requirements are met. If the requirements are not met and the maximum number of iterations is not exceeded, then parameter feedback adjustments will be made. Otherwise, output the final prediction result; Parameter feedback adjustments include: The prediction error information is fed back to the weight-aware fractional adaptive genetic algorithm WAFA-GA for parameter optimization, resulting in the third observation matrix, which includes: Based on the prediction error information, adjust the range of values ​​for the fractional-order parameter. When the prediction error is greater than the set threshold, increase the fractional-order parameter to speed up the parameter adjustment. When the prediction error is less than the set threshold, decrease the value to improve convergence stability. Based on the time-series distribution characteristics of the error, the decay coefficient is dynamically adjusted to optimize the decay characteristics of the Mittag-Leffler function, thereby achieving adaptive adjustment of crossover probability and mutation probability. Based on the prediction accuracy, the weight coefficients in the multi-objective fitness function are recalculated to balance the importance weights of multiple optimization objectives in the first observation matrix optimization process. The predicted power of the wind turbine is obtained based on the third observation matrix.

Images