An improved-lstm-based overall state evaluation model for a wind power device

Abstract

The application discloses a kind of wind power equipment whole state evaluation model based on improved LSTM, this paper proposes LSTM-SOM-KELM model, the function of long short-term memory network (LSTM) forget gate and input gate is fused into the method of dual gated LSTM (Dual Gated LSTM, DG-LSTM) in update gate, through the synergistic effect of update gate and reset gate, quickly adapt to the short-term fluctuations of wind power equipment state, while reducing model complexity, reduce the risk of overfitting.DG-LSTM auto-encoder extraction component level residual feature is constructed, self-organizing mapping (SOM) is used to fuse component residual, and sparrow search algorithm (SSA) is used to optimize kernel extreme learning machine (KELM) parameter, to establish health index calculation model.The experimental results show that DG-LSTM model is better than traditional LSTM and GRU model in training effect, verification performance and prediction accuracy;The mean absolute error (MAE) and mean square error (MSE) of SSA-KELM are reduced by 0.008 and 0.128 respectively, which significantly improves the accuracy and reliability of wind power equipment state evaluation.

Description

Technical Field

[0001] This invention relates to the field of LSTM wind power equipment technology, and specifically to a wind power equipment overall condition assessment model based on an improved LSTM. Background Technology

[0002] In wind power equipment condition assessment tasks, the equipment operating cycle is typically long, and its health state evolution exhibits significant long-term dependencies. Therefore, the model needs to be able to capture long-term temporal dependencies. LSTM, with its cellular state mechanism and gating structure, performs excellently in processing long-sequence information and serves as the foundational structure for this study.

[0003] Reference [1] uses Long Short Term Memory (LSTM) to predict wind power, and its prediction performance is better than that of traditional machine learning algorithms. Reference [2] proposes an enhanced forget gate LSTM model for time series prediction. Compared with the standard LSTM, the prediction accuracy is significantly improved. Reference [3] embeds wavelets into LSTM, Reference [4] uses a heap optimizer to train LSTM, Reference [5] uses a genetic algorithm to optimize the LSTM layers, and Reference [6] uses an improved sparrow algorithm to adjust the LSTM model parameters, etc. However, wind turbines are not always in the process of fault evolution, and there are a lot of short-term fluctuations in their operating state. These short-term characteristics are often of great significance for the early identification of faults. When dealing with such short-term changes, traditional LSTM may cause response lag or over-smoothing due to the long information flow path. Compared with LSTM, GRU has a relatively simple structure with only two gated units, fewer parameters, and lower computational complexity. Reference [7] proposed a state monitoring method based on the spatiotemporal fusion of convolutional neural networks (CNN) and gated recurrent units (GRU). A deep learning model is constructed using CNN and GRU. After extracting and fusing the spatiotemporal features in the SCADA system, the model is trained. The health status of the wind turbine is identified based on the residual between the actual data and the predicted data. It is suitable for short-term dependencies, but it is not sufficient for processing information with longer interval dependencies, thus affecting the model's comprehensive and accurate assessment of the wind power equipment status.

[0004] In the field of wind power equipment condition assessment, random forests and SVMs are widely used due to their respective advantages, but they also have some limitations. Random forests perform well in handling high-dimensional data and feature selection, but their training and prediction processes are computationally complex and time-consuming. Furthermore, they are sensitive to noisy data, which can lead to overfitting, and due to their randomness, the results are difficult to completely replicate. SVMs perform well on small datasets and in high-dimensional spaces, and can handle linear and non-linear classification problems, but they are computationally expensive on large datasets and cannot directly handle multi-class classification problems, requiring expansion strategies. To overcome these shortcomings, the KELM (Kernel Extreme Learning Machine) model was developed. KELM is an extreme learning machine algorithm based on kernel methods. It maps input data to a high-dimensional feature space through kernel tricks and performs linear regression in the feature space. KELM is based on a single hidden layer feedforward neural network, achieving linear separability by randomly initializing input weights and kernel functions to map data to a high-dimensional space. It has the advantages of fast training speed and suitability for small sample sizes and non-linear data. However, KELM is sensitive to the choice of kernel function and has low sensitivity to mixed fault characteristics. To address these shortcomings, this paper improves KELM using SSA to further enhance the model's performance and applicability.

[0005] To address this, this paper introduces the update and reset gate mechanisms from GRU to improve the standard LSTM. The update gate simplifies the gating structure by incorporating the functions of the forget gate and the input gate, while enhancing the model's ability to balance long-term and short-term information. The reset gate allows the model to dynamically ignore some historical information, thus quickly adapting to short-term anomalies. This hybrid gating mechanism enables the improved LSTM to maintain its long-term dependency modeling capabilities while more sensitively capturing short-term state fluctuations, thereby improving the accuracy and robustness of wind power equipment condition assessment.

[0006] In summary, this paper proposes an LSTM-SOM-KELM model. By introducing the update and reset gate mechanisms of GRU into wind turbine condition assessment, the dual-gated LSTM (DG-LSTM) enhances the model's sensitivity and adaptability to short-term information while retaining the long-term dependency capture capability of LSTM. This allows for a more comprehensive and accurate assessment of wind turbine condition, including the grasp of long-term trends and timely response to short-term anomalies. Furthermore, this paper constructs a component-level residual feature extraction method based on a DG-LSTM autoencoder and combines it with self-organizing maps (SOM) for component residual fusion. Simultaneously, the Sparrow Search Algorithm (SSA) is used to optimize the parameters of the Kernel Extreme Learning Machine (KELM) to establish a health index calculation model. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention is achieved through the following technical solution: a wind power equipment overall condition assessment model based on an improved LSTM, which obtains a wind farm power generation dataset and preprocesses the wind farm power generation dataset to generate a wind power preprocessed dataset; the preprocessing includes: missing value processing, outlier processing, and normalization processing; A TPA-TCN network is established based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network. Based on the TPA-TCN network, features are extracted from the wind power preprocessing dataset to generate a wind power feature dataset. The Long Short-Term Memory network is cascaded with the TPA-TCN network to construct the TPA-TCN-LSTM network model; The hyperparameters in the TPA-TCN-LSTM network model are optimized and iterated using the Sparrow Search algorithm to generate a TPA-TCN-LSTM network model with optimal parameters for wind power prediction.

[0008] The wind power feature dataset is divided into a training set and a test set; The TPA-TCN-LSTM network model with the optimal parameters is trained based on the training set. The TPA-TCN-LSTM network model with the optimal parameters after training is validated using the test set.

[0009] As a further optimization of the invention, the wind farm power generation dataset is preprocessed to generate wind power preprocessed data, including: The wind farm power generation dataset is subjected to outlier detection using a density-based clustering method with noise, and outlier data is filtered out. The outlier data in the wind farm power generation dataset is deleted to generate an outlier-free dataset. Based on the KNN algorithm, the nearest neighbor samples of the outlier data are selected to estimate and fill the gaps in the outlier-free dataset, generating an outlier-free filled dataset. The abnormal-free filled dataset is normalized to generate a wind power preprocessing dataset.

[0010] As a further optimization of the invention, a TPA-TCN network is established based on temporal convolutional networks and temporal pattern attention mechanisms, including: The wind power time series in the wind power preprocessing dataset is transformed into a supervised learning problem sequence. Construct a temporal convolutional network, which includes: one-dimensional full convolution, causal convolution, residual connection module and dilated convolution module; Based on the temporal convolutional network, the temporal pattern attention mechanism, and the supervised learning problem sequence, a TPA-TCN network is constructed.

[0011] As a further optimization of the invention, the Long Short-Term Memory network includes an input gate, a forget gate, and an output gate; The forget gate is used to determine the data to be discarded from the cells of the long short-term memory network; The input gate is used to determine the data used to update the memory state from the input wind power feature dataset; The output gate is used to determine the output content based on the input wind power feature dataset and the memory of the cells of the long short-term memory network.

[0012] Concatenating a Long Short-Term Memory (LSTM) network with a TPA-TCN network to construct a TPA-TCN-LSTM network model involves training the forward propagation and the back propagation using an LSTM neural network.

[0013] As a further optimization of the invention, the forward propagation for training the LSTM neural network includes: setting the parameter data of the LSTM neural network; Calculate the nonlinear activation function of the hidden layer based on the parameter data; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input of the input gate, the output of the input gate, the input of the forget gate, and the output of the forget gate; Based on the parameter data, the output of the input gate, and the output of the forget gate, calculate the input of the storage unit and the state value of the storage unit; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input and output of the output gate; The unit output is calculated based on the state value of the storage unit and the output of the output gate; wherein the unit output is the prediction result of the TPA-TCN-LSTM network model.

[0014] As a further optimization of the invention, the backpropagation for training the LSTM neural network includes: The loss function is determined based on the unit output and the actual value. Calculate the weight correction term based on the loss function and the parameter data; Based on the unit output and the loss function, define the partial derivative of the loss function with respect to the storage unit output; Based on the state value of the storage cell and the loss function, define the partial derivative of the loss function with respect to the state of the storage cell; Based on the parameter data, calculate the partial derivative of the loss function with respect to the storage cell; The output gate weight correction term is calculated based on the input of the output gate, the state value of the storage cell, and the partial derivative of the loss function with respect to the output of the storage cell. Calculate the weight correction term of the storage cell based on the parameter data and the partial derivative of the loss function with respect to the storage cell; The forget gate weight correction term is calculated based on the input of the forget gate and the partial derivative of the loss function with respect to the state of the storage cell. The input gate weight correction term is calculated based on the input of the input gate, the input of the storage cell, and the partial derivative of the loss function with respect to the state of the storage cell.

[0015] As a further optimization of the invention, the hyperparameters in the TPA-TCN-LSTM network model are optimized iteratively according to the sparrow search algorithm to generate a TPA-TCN-LSTM network model with optimal parameters, including: The parameters of the sparrow search algorithm are initialized based on the hyperparameters in the TPA-TCN-LSTM network model. Generate a sparrow position matrix based on the parameters; Based on the sparrow search algorithm and the sparrow position matrix, a fitness value matrix for sparrows is generated; Update the positions of producers and beggars until the optimal position and fitness value of the sparrow are obtained; The optimal position and optimal fitness value are assigned to the TPA-TCN-LSTM network model as optimal parameters to generate a TPA-TCN-LSTM network model with optimal parameters.

[0016] As a further optimization of the invention, the system includes: a data acquisition module, a data processing module, an establishment module, an extraction module, an optimization iteration module, and a prediction module; The data acquisition module is used to acquire a wind farm power generation dataset. The data processing module is used to preprocess the wind farm power generation dataset to generate a wind power preprocessed dataset; the preprocessing includes: missing value processing, outlier processing and normalization processing; The establishment module is used to establish a TPA-TCN network based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network; The extraction module is used to extract features from the wind power preprocessing dataset based on the TPA-TCN network to generate a wind power feature dataset. The optimization iteration module is used to optimize and iterate the hyperparameters in the TPA-TCN-LSTM network model according to the sparrow search algorithm, and generate the TPA-TCN-LSTM network model with optimal parameters. The prediction module is used to predict wind power using the TPA-TCN-LSTM network model with the optimal parameters. Beneficial effects

[0017] This invention provides a wind turbine overall condition assessment model based on an improved LSTM, which has the following advantages: This paper addresses the issues of information imbalance between long-term and short-term data and delayed response to sudden faults in wind power equipment condition assessment. A hybrid assessment model based on an improved LSTM, LSTM-SOM-KELM, is proposed. A DG-LSTM structure is introduced, integrating the gating mechanisms of LSTM and GRU, significantly improving the response speed and adaptability to short-term state changes while retaining long-term memory capabilities. A feature fusion module based on SOM is constructed to effectively integrate residual features from multiple components, enhancing the integrity and interpretability of state information. SSA optimization of KELM parameters overcomes the sensitivity of traditional KELM to kernel functions, improving the accuracy and robustness of health status assessment. Experimental results show that the proposed model outperforms traditional LSTM, GRU, SVM, and random forest models in terms of training efficiency, validation performance, and prediction accuracy. DG-LSTM shows the best performance across multiple metrics, while SSA-KELM reduces MAE and MSE by 0.008 and 0.128, respectively, verifying the model's significant effect on improving the timeliness and reliability of condition assessment. Attached Figure Description

[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a wind power equipment condition assessment model based on an improved LSTM. Figure 2 This is a schematic diagram of an improved process for obtaining a dataset of wind farm power generation.

[0019] Figure 3 This is a schematic diagram of the process structure of a temporal convolutional network with an improved temporal attention mechanism.

[0020] Figure 4 This is a schematic diagram of the process structure of an improved sparrow search algorithm.

[0021] Figure 5 A schematic diagram illustrating the steps of a TCN-LSTM-based wind power equipment condition assessment model.

[0022] Figure 6 This is a diagram of the DG-LSTM model structure.

[0023] Figure 7 The network structure is 6LSTM-SOM-KELM model.

[0024] Figure 8 This is a comparison of the predicted results of the environmental average wind speed monitoring index.

[0025] Figure 9 This is a comparison of the predicted results of the environmental maximum wind speed monitoring index.

[0026] Figure 10 This is a comparison of the predicted results of the minimum wind speed monitoring index for the environment.

[0027] Figure 11 This is a comparison of the predicted results of the generator's maximum speed monitoring index.

[0028] Figure 12 This is a comparison of the predicted results of the generator average speed monitoring index.

[0029] Figure 13 This is a comparison of the predicted results of the generator minimum speed monitoring index.

[0030] Figure 14 This is a comparison of the predicted results of the average rotational speed monitoring index.

[0031] Figure 15 This is a comparison of the predicted results of the maximum rotational speed monitoring index.

[0032] Figure 16 This is a comparison of the predicted results for the minimum rotational speed monitoring index. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments. Example

[0034] Please see Figures 1-6 , This invention provides a wind power equipment condition assessment model based on an improved LSTM. A wind power equipment overall condition assessment model based on an improved LSTM is provided, which obtains a wind farm power generation dataset and preprocesses the wind farm power generation dataset to generate a wind power preprocessed dataset; the preprocessing includes: missing value processing, outlier processing and normalization processing. A TPA-TCN network 12 is established based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network. Based on the TPA-TCN network, features are extracted from the wind power preprocessing dataset to generate wind power feature dataset 13; The Long Short-Term Memory network is cascaded with the TPA-TCN network to construct the TPA-TCN-LSTM network model 14; The hyperparameters in the TPA-TCN-LSTM network model are optimized and iterated using the Sparrow Search algorithm to generate the optimal TPA-TCN-LSTM network model for wind power prediction.15

[0035] The wind power feature dataset is divided into a training set and a test set 16; The TPA-TCN-LSTM network model with the optimal parameters is trained based on the training set 17; The TPA-TCN-LSTM network model with the optimal parameters after training was validated using the test set.18

[0036] The wind farm power generation dataset is preprocessed to generate wind power preprocessed data 11, including: The wind farm power generation dataset was subjected to outlier detection using a density-based clustering method with noise, and outlier data 111 were filtered out. The outlier data in the wind farm power generation dataset is deleted to generate an outlier-free dataset 112. Based on the KNN algorithm, the nearest neighbor samples of the outlier data are selected to estimate and fill the gaps in the outlier-free dataset, generating an outlier-free filled dataset 113. The abnormal-free data set is normalized to generate wind power preprocessing data set 114.

[0037] A TPA-TCN network is established based on temporal convolutional networks and temporal pattern attention mechanisms, including: The wind power time series in the wind power preprocessing dataset is transformed into a supervised learning problem sequence 121; Construct a temporal convolutional network, which includes: a one-dimensional fully convolutional network, a causal convolutional network, a residual connection module, and a dilated convolutional module 122; Based on the temporal convolutional network, the temporal pattern attention mechanism, and the supervised learning problem sequence, a TPA-TCN network 123 is constructed.

[0038] The long short-term memory network includes an input gate, a forget gate, and an output gate; The forget gate is used to determine the data to be discarded from the cells of the long short-term memory network; The input gate is used to determine the data used to update the memory state from the input wind power feature dataset; The output gate is used to determine the output content based on the input wind power feature dataset and the memory of the cells in the long short-term memory network.

[0039] Concatenating a Long Short-Term Memory (LSTM) network with a TPA-TCN network to construct a TPA-TCN-LSTM network model involves training the forward propagation and the back propagation using an LSTM neural network.

[0040] The forward propagation for training the LSTM neural network includes: setting the parameter data of the LSTM neural network; Calculate the nonlinear activation function of the hidden layer based on the parameter data; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input of the input gate, the output of the input gate, the input of the forget gate, and the output of the forget gate; Based on the parameter data, the output of the input gate, and the output of the forget gate, calculate the input of the storage unit and the state value of the storage unit; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input and output of the output gate; The unit output is calculated based on the state value of the storage unit and the output of the output gate; wherein the unit output is the prediction result of the TPA-TCN-LSTM network model.

[0041] The backpropagation training of the LSTM neural network includes: The loss function is determined based on the unit output and the actual value. Calculate the weight correction term based on the loss function and the parameter data; Based on the unit output and the loss function, define the partial derivative of the loss function with respect to the storage unit output; Based on the state value of the storage cell and the loss function, define the partial derivative of the loss function with respect to the state of the storage cell; Based on the parameter data, calculate the partial derivative of the loss function with respect to the storage cell; The output gate weight correction term is calculated based on the input of the output gate, the state value of the storage cell, and the partial derivative of the loss function with respect to the output of the storage cell. Calculate the weight correction term of the storage cell based on the parameter data and the partial derivative of the loss function with respect to the storage cell; The forget gate weight correction term is calculated based on the input of the forget gate and the partial derivative of the loss function with respect to the state of the storage cell. The input gate weight correction term is calculated based on the input of the input gate, the input of the storage cell, and the partial derivative of the loss function with respect to the state of the storage cell.

[0042] The hyperparameters in the TPA-TCN-LSTM network model are optimized iteratively using the sparrow search algorithm to generate a TPA-TCN-LSTM network model 15 with optimal parameters, including: The parameters 151 of the sparrow search algorithm are initialized according to the hyperparameters in the TPA-TCN-LSTM network model. Based on the parameters, a sparrow position matrix 152 is generated; Based on the sparrow search algorithm and the sparrow position matrix, a fitness value matrix 153 for sparrows is generated; Update the positions of producers and beggars until the optimal position and fitness value of 154 for the sparrow are obtained; The optimal position and optimal fitness value are assigned to the TPA-TCN-LSTM network model as optimal parameters to generate the TPA-TCN-LSTM network model 155 with optimal parameters.

[0043] The system includes: a data acquisition module, a data processing module, an establishment module, an extraction module, an optimization iteration module, and a prediction module; The data acquisition module is used to acquire a wind farm power generation dataset. The data processing module is used to preprocess the wind farm power generation dataset to generate a wind power preprocessed dataset; the preprocessing includes: missing value processing, outlier processing and normalization processing; The establishment module is used to establish a TPA-TCN network based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network; The extraction module is used to extract features from the wind power preprocessing dataset based on the TPA-TCN network to generate a wind power feature dataset. The optimization iteration module is used to optimize and iterate the hyperparameters in the TPA-TCN-LSTM network model according to the sparrow search algorithm, and generate the TPA-TCN-LSTM network model with optimal parameters. The prediction module is used to predict wind power using the TPA-TCN-LSTM network model with the optimal parameters.

[0044] The working principle of the above technical solution is explained below: When using this invention, 1.1 LSTM Algorithm An LSTM (Laser-Based Memory) consists of an input gate, a forget gate, an output gate, and a cell state. The forget gate determines the degree to which the cell state information from the previous time step is forgotten; the input gate controls the degree to which the current input information is written into the cell state; and the output gate adjusts the output of the currently hidden state. The update formula for an LSTM cell is: Forgotten Gate: ; Input Gate: ; Candidate memory units: ; Output gate: ; Memory unit update: ; Hidden state: ; The forget gate and input gate independently filter and control information, lacking a unified mechanism to integrate and coordinate the flow of information from these two parts. This may lead to situations where the model cannot accurately balance the retention of past information and the absorption of current information. When the wind power equipment is operating stably, the model excessively retains past information, while when abnormal short-term fluctuations occur, it cannot quickly adjust and update its assessment of the equipment status in a timely manner.

[0045] 1.2 GRU Algorithm GRU integrates the forget gate and input gate functions into an update gate, and introduces a reset gate to control the influence of the previous hidden state on the candidate hidden state. The update formula for a GRU cell is: Update Gate: ; Reset Door: ; Candidate hidden state: ; Hidden state: ; While the update gate of GRU can effectively control the degree of information fusion between the hidden state of the previous time step and the candidate hidden state of the current time step, it lacks a long-term memory management mechanism like the cell state in LSTM. In tasks such as wind power equipment condition assessment that require processing long-term series data, it may not be able to effectively manage and utilize long-term dependent information, and is easily disturbed by short-term fluctuations, resulting in the loss of important long-term trend information.

[0046] 2.1 DG-LSTM (I) Implementation of the update gate mechanism in LSTM 1. Update the door design In LSTM, an update gate is introduced, and its calculation formula is as follows:

[0047] This indicates updating the gate output, where σ is the sigmoid activation function, which restricts the output value to the range (0,1). It updates the gate weight matrix. It is a bias term. The state was hidden in the previous moment. Enter the current time.

[0048] 2. Updates to cell state and hidden state Drawing inspiration from the GRU update mechanism, the improved LSTM cell state update formula is as follows: ; Meanwhile, the update formula for the hidden state is: ; It represents the current state of the candidate cells. The candidate hidden state is calculated based on the candidate cell state. (Update gate) It determines the fusion ratio of the previous state information and the current candidate state information. The closer the value is to 1, the more information from the previous time step is retained; the closer the value is to 0, the more it depends on the current candidate state.

[0049] (II) Implementation of the reset gate mechanism in LSTM 1. Reset door design The formula for adding a reset door is: ; It resets the gate output. It is the reset gate weight matrix. It is a bias term.

[0050] 2. Generation of candidate hidden states When generating candidate hidden states, the hidden states from the previous time step are weighted by a reset gate: ; Where ϕ is typically the tanh activation function, W is the weight matrix, and b is the bias term. Reset gate The influence of the previous hidden state on the current candidate hidden state is controlled. When the value is close to 0, the interference of the previous information is reduced, and the candidate hidden state depends more on the current input.

[0051] The structure diagram of the DG-LSTM algorithm is as follows: Figure 6 As shown: By integrating the forget gate and input gate functions of LSTM into the update gate and introducing a reset gate mechanism, not only is the long-term dependency capture capability of LSTM preserved, but it can also more efficiently filter and process short-term fluctuation information in wind power equipment operation data, improving the model's sensitivity and adaptability to changes in wind power equipment status; the number of parameters is reduced, lowering model complexity and overfitting risk; and the reset gate is used to dynamically control information flow, enabling the model to better handle short-term and long-term information, enhancing its adaptability and flexibility to changes in input sequence.

[0052] 2.2 SOM Self-organizing map (SOM) is an unsupervised neural network algorithm used for feature dimensionality reduction and cluster analysis. In the feature fusion stage, the SOM algorithm is used to reduce the dimensionality of the extracted features while simultaneously performing cluster analysis to obtain the overall health characteristics of the wind turbine.

[0053] SOM formula: ; in, It is the neuron weight. It's the learning rate. It is a neighborhood function. It is the input vector.

[0054] 2.3 SSA-KELM This study uses an optimized KELM model to train a health status assessment model.

[0055] The output formula of the KELM model is: ; Wherein K(X) i The kernel function is X, which is usually a radial basis function (RBF). ; To optimize the parameters of the KELM model, the Sparrow Search Algorithm (SSA) is adopted. SSA is a swarm intelligence optimization algorithm based on sparrow foraging behavior, which has strong global search capabilities and fast convergence speed.

[0056] The main steps of the SSA algorithm are as follows: Initialize the location of the sparrow population, and randomly generate the location vector X of N individual sparrows. i Each position vector represents a set of parameters of the KELM model.

[0057] Calculate the fitness value for each individual sparrow. The fitness value is defined as the mean squared error (MSE) of the KELM model on the training set.

[0058] The sparrow population was sorted according to its fitness value, and the sparrow with the smallest fitness value (N) was selected. p One sparrow acts as the producer, while the rest of the sparrows act as followers.

[0059] Producer update rule: Individual producers randomly search for food sources in the search space and update their position vectors accordingly. ; Where δ is the step size factor, L is the Lévy flight step size, β is the decay coefficient, and t is the current iteration number.

[0060] Follower update rule: Follower individuals move towards the position of producer individuals and update their position vectors. ; in, This is the average position vector of individual producers.

[0061] Watchdog Update Rules: A sparrow is randomly selected as the watchdog. When a predator is detected, the watchdog sounds an alarm, and all sparrows move to a safe area. ; in, This is the position vector of a randomly selected individual sparrow.

[0062] Repeat the above steps until the maximum number of iterations is reached or the stopping condition is met.

[0063] The network structure of the LSTM-SOM-KELM model is as follows: Figure 7 As shown.

[0064] This paper uses raw data from a wind farm SCADA system publicly available on EDP. The wind turbine has a rated power of 2,000 kW, and the cut-in wind speed, rated wind speed, and cut-out wind speed are 4, 12, and 25 m / s, respectively. The dataset contains 80 wind turbine features, including ambient wind speed, power, and bearing temperature, with a sampling interval of 10 min.

[0065] Data from wind turbine No. 1, after normalization of feature variables selected through feature selection, was used as training data to construct the standard model, while data from wind turbine No. 6 was used as test data. 3.1 Comparative Analysis of LSTM, GRU, and DG-LSTM Models (I) Comparison of evaluation indicators The experiment compared three models: traditional LSTM, GRU, and DG-LSTM. Five metrics were compared: training loss, validation loss, mean health index, mean absolute error (MAE), and mean squared error (MSE). Training loss and validation loss reflect the model's fit during training; the mean health index represents the model's predictive performance, with higher values ​​indicating closer predictions to the true values; MAE and MSE are common prediction error metrics, measuring the mean absolute difference and squared difference between predicted and true values, respectively, with smaller values ​​indicating higher prediction accuracy. The comparison results of the autoencoder models are shown in Table 1. Table 1 Comparison of Autoencoder Models Module Tr_Loss Val_Loss M_H_I MAE MSE LSTM 0.285 0.250 1.289 0.172 0.093 GRU 0.267 0.235 1.292 0.174 0.079 DG-LSTM 0.254 0.219 1.346 0.143 0.066 In terms of training loss, LSTM has a training loss of 0.2855, GRU has 0.2679, while DG-LSTM's loss drops to 0.2547. This indicates that DG-LSTM fits the data better on the training set, reducing the training loss by approximately 4.5% compared to LSTM. Regarding validation loss, LSTM has 0.2501, GRU has 0.2360, and DG-LSTM has 0.2198. DG-LSTM has the lowest validation loss, indicating that it maintains its ability to fit the training set while also exhibiting the strongest generalization ability, improving validation performance by approximately 6.1%. This suggests that DG-LSTM's gating mechanism optimizes information flow, reduces overfitting, and makes it perform better on unseen data.

[0066] The average health index of LSTM is 1.2895, GRU is slightly higher at 1.2928, while DG-LSTM reaches 1.3461. This indicates that DG-LSTM's predicted values ​​are more consistent with the actual health status in the health status prediction task. Its improved gating structure can more accurately capture the health information features in the data, making the prediction results closer to the true values.

[0067] In terms of prediction accuracy (MAE), LSTM achieved 0.1722, GRU 0.1745, and DG-LSTM reduced to 0.1433. In terms of prediction precision (MSE), DG-LSTM significantly outperformed the other two models, reducing MAE by approximately 17% and MSE by approximately 28% compared to LSTM. This further confirms the advantage of DG-LSTM in prediction accuracy, demonstrating its more effective fusion processing of long-term information from sequence data and its ability to more accurately predict target values.

[0068] (II) Comparison of Actual Values ​​and Predicted Values Figures 8 to 16 To compare the prediction results of different models for each monitoring indicator, the horizontal axis represents the sampling points, and the vertical axis represents the predicted value of each indicator. The value for March 1, 2016 is selected.

[0069] As shown in the figure, both the LSTM and GRU models exhibit certain errors during the prediction process, especially when data fluctuations are large or abrupt changes occur, resulting in significant deviations between the predicted and actual curves. For example, in predicting the ambient average wind speed (Amb_WindSpeed_Avg), the predicted values ​​of both LSTM and GRU failed to accurately follow the fluctuations of the actual value at multiple sampling points, showing lag or significant deviations. In contrast, the improved LSTM model shows a higher degree of fit between the predicted and actual curves, more accurately capturing data trends and demonstrating superior fitting ability and prediction accuracy when handling complex fluctuations, effectively reducing prediction errors. This verifies the positive effect of introducing update and reset gate mechanisms on improving the performance of the LSTM model, making it more suitable for the need for accurate data prediction in wind power equipment condition assessment.

[0070] 3.2 Analysis of Predictive Comparative Experiment Results This paper selects mean absolute error (MAE) and mean squared error (MSE) as indicators to evaluate the model, and the prediction results are shown in Table 2.

[0071] Table 2 Comparison of Health Index Prediction Models Module MAE MSE SSA-KELM 1.293 2.064 KELM 1.301 2.192 SVM 1.306 2.452 Random Forest 1.337 3.010 The SSA-KELM model performed best among all models, with the lowest mean absolute error (MAE) and mean squared error (MSE) at 1.293 and 2.064, respectively. This indicates that SSA-KELM has higher accuracy in prediction compared to other models, capturing data features more precisely and providing more reliable prediction results. The KELM model was second best; the SVM model had a slightly higher error value (MAE and MSE of 1.306 and 2.452, respectively), indicating slightly lower prediction accuracy. The RandomForest model had the highest error value (MAE and MSE of 1.337 and 3.010, respectively). This may be because RandomForest has a relatively weaker ability to capture data features, or it may have limitations in handling complex problems such as wind power equipment health status prediction.

[0072] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0073] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A wind power equipment overall condition assessment model based on improved LSTM, which obtains wind farm power generation data set and preprocesses the wind farm power generation data set (11) to generate wind power preprocessing data set; The preprocessing includes: missing value processing, outlier processing, and normalization processing; A TPA-TCN network (12) is established based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network; Based on the TPA-TCN network, the wind power preprocessing dataset is used to extract features and generate a wind power feature dataset (13). The long short-term memory network is cascaded with the TPA-TCN network to construct the TPA-TCN-LSTM network model (14). The hyperparameters in the TPA-TCN-LSTM network model are optimized and iterated according to the sparrow search algorithm to generate the optimal TPA-TCN-LSTM network model for wind power prediction (15). The wind power feature dataset is divided into a training set and a test set (16). The TPA-TCN-LSTM network model with the optimal parameters is trained based on the training set (17). The TPA-TCN-LSTM network model with the optimal parameters after training was validated based on the test set (18).

2. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: The wind farm power generation dataset is preprocessed to generate wind power preprocessed data (11), including: The wind farm power generation dataset is subjected to outlier detection using a density-based clustering method with noise, and outlier data is filtered out (111). The outlier data in the wind farm power generation dataset is deleted to generate an outlier-free dataset (112). The nearest neighbor samples of the outlier data are selected according to the KNN algorithm to estimate and fill the gaps in the outlier-free dataset, generating an outlier-free filled dataset (113). The abnormal-free data set is normalized to generate a wind power preprocessing data set (114).

3. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: A TPA-TCN network (12) is established based on temporal convolutional networks and temporal pattern attention mechanisms, including: Transform the wind power time series in the wind power preprocessing dataset into a supervised learning problem sequence (121); Construct a temporal convolutional network, which includes: one-dimensional full convolution, causal convolution, residual connection module and dilated convolution module (122). Based on the temporal convolutional network, the temporal pattern attention mechanism, and the supervised learning problem sequence, a TPA-TCN network (123) is constructed.

4. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: The long short-term memory network includes an input gate, a forget gate, and an output gate; The forget gate is used to determine the data to be discarded from the cells of the long short-term memory network; The input gate is used to determine the data used to update the memory state from the input wind power feature dataset; The output gate is used to determine the output content based on the input wind power feature dataset and the memory of the cells of the long short-term memory network; Concatenating a Long Short-Term Memory (LSTM) network with a TPA-TCN network to construct a TPA-TCN-LSTM network model involves training the forward propagation and the back propagation using an LSTM neural network.

5. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: The forward propagation for training the LSTM neural network includes: setting the parameter data of the LSTM neural network; Calculate the nonlinear activation function of the hidden layer based on the parameter data; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input of the input gate, the output of the input gate, the input of the forget gate, and the output of the forget gate; Based on the parameter data, the output of the input gate, and the output of the forget gate, calculate the input of the storage unit and the state value of the storage unit; Based on the parameter data and the nonlinear activation function of the hidden layer, calculate the input and output of the output gate; The unit output is calculated based on the state value of the storage unit and the output of the output gate; wherein the unit output is the prediction result of the TPA-TCN-LSTM network model.

6. A wind turbine overall condition assessment model based on an improved LSTM as described in claim 1 (the reference needs modification), characterized in that: The backpropagation training of the LSTM neural network includes: The loss function is determined based on the unit output and the actual value. Calculate the weight correction term based on the loss function and the parameter data; Based on the unit output and the loss function, define the partial derivative of the loss function with respect to the storage unit output; Based on the state value of the storage cell and the loss function, define the partial derivative of the loss function with respect to the state of the storage cell; Based on the parameter data, calculate the partial derivative of the loss function with respect to the storage cell; The output gate weight correction term is calculated based on the input of the output gate, the state value of the storage cell, and the partial derivative of the loss function with respect to the output of the storage cell. Calculate the weight correction term of the storage cell based on the parameter data and the partial derivative of the loss function with respect to the storage cell; The forget gate weight correction term is calculated based on the input of the forget gate and the partial derivative of the loss function with respect to the state of the storage cell. The input gate weight correction term is calculated based on the input of the input gate, the input of the storage cell, and the partial derivative of the loss function with respect to the state of the storage cell.

7. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: The hyperparameters in the TPA-TCN-LSTM network model are optimized and iterated according to the sparrow search algorithm to generate the optimal TPA-TCN-LSTM network model (15), including: The parameters of the sparrow search algorithm are initialized according to the hyperparameters in the TPA-TCN-LSTM network model (151). Based on the parameters, generate the sparrow position matrix (152). Based on the sparrow search algorithm and the sparrow position matrix, a fitness value matrix (153) of sparrows is generated. Update the positions of producers and beggars until the optimal position and fitness value of the sparrow are obtained (154). The optimal position and optimal fitness value are assigned to the TPA-TCN-LSTM network model as optimal parameters to generate a TPA-TCN-LSTM network model with optimal parameters (155).

8. The wind turbine overall condition assessment model based on improved LSTM according to claim 1, characterized in that: The system includes: a data acquisition module, a data processing module, an establishment module, an extraction module, an optimization iteration module, and a prediction module; The data acquisition module is used to acquire a wind farm power generation dataset. The data processing module is used to preprocess the wind farm power generation dataset to generate a wind power preprocessed dataset; the preprocessing includes: missing value processing, outlier processing and normalization processing; The establishment module is used to establish a TPA-TCN network based on a temporal convolutional network and a temporal pattern attention mechanism; wherein, the temporal pattern attention mechanism is used to improve the performance of the temporal convolutional network; The extraction module is used to extract features from the wind power preprocessing dataset based on the TPA-TCN network to generate a wind power feature dataset. The optimization iteration module is used to optimize and iterate the hyperparameters in the TPA-TCN-LSTM network model according to the sparrow search algorithm, and generate the TPA-TCN-LSTM network model with optimal parameters. The prediction module is used to predict wind power using the TPA-TCN-LSTM network model with the optimal parameters.

Images