Temperature early warning method for slip ring chamber of hydropower station based on fusion of PCA and hedged GRU optimized by PSO

Abstract

The present invention discloses a temperature early warning method for the slip ring chamber of a hydropower station based on the fusion of PCA and GRU optimized by PSO. The method takes temperature-related data as input, uses a variance elimination method to delete irrelevant data, uses a particle swarm algorithm to optimize the PCA dimensionality reduction algorithm, finds the optimal dimensionality reduction dimension, uses an attention mechanism to adjust the weights of the reduced data, uses the weighted data as input of a gated recurrent unit (GRU), introduces a hedging layer, a residual connection and other structures to improve the GRU, and finally uses the residual value as the alarm threshold. The method solves the problems in the prior art that the method is easily affected by the threshold setting, the model extracts incomplete and inaccurate features, the model is single and the early warning effect is poor.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical fiber fault detection, and in particular relates to a temperature early warning method for a slip ring chamber of a hydropower station based on the fusion of PCA optimized by PSO and GRU. Background Art

[0002] Long-term high temperatures in the slip ring chamber of a generator set can accelerate the aging of the insulation in the slip ring chamber, shortening the service life of the equipment. In severe cases, this can cause the slip rings to spark, the excitation system to short-circuit or ground, and other serious consequences such as electrical accidents and shutdowns. To address the problem of excessively high slip ring chamber temperatures and prevent aging and breakdown of the insulation in the slip ring chamber, which can cause unexpected shutdowns, it is necessary to provide a temperature warning for the slip ring chamber. Existing technologies generally use fixed threshold methods, machine learning methods, and deep learning methods for temperature warnings. These technologies have the following disadvantages:

[0003] The fixed threshold method determines whether a temperature alarm occurs by setting a temperature threshold. Although this method is simple to operate, it is easily limited by the absolute limit of the threshold, and the setting of the threshold is easily affected by human experience.

[0004] Machine learning methods collect temperature-related data and use this data to build a fault warning model to provide temperature warnings. While this method is superior to fixed threshold methods, it suffers from issues such as incomplete and inaccurate features extracted by the model, resulting in poor warning effectiveness.

[0005] The deep learning method, similar to the machine learning method, performs temperature warning by establishing a deep learning temperature warning model. This method does not further process the input data and is prone to extracting features of irrelevant data. It also has disadvantages such as a single model.

[0006] The equipment in the slip ring room of a hydropower station will be affected by multiple parameters such as load, temperature, and external environment during long-term operation, so multiple parameters need to be monitored. In order to reduce the impact of irrelevant parameters on the prediction results, data dimensionality reduction is required. In order to ensure that the extracted features are more complete, more original data information needs to be retained while reducing the dimensionality. Since each parameter has a different contribution rate to the prediction results, the corresponding weight of each parameter needs to be adjusted according to the corresponding contribution rate. Extracting data features more comprehensively and accurately is also an effective means to improve the effect of temperature warning.

[0007] Therefore, it is necessary to design a temperature warning method for the slip ring chamber of a hydropower station based on the fusion of PCA optimized by PSO and hedging GRU to solve the above problems. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a temperature warning method for the slip ring chamber of a hydropower station based on the fusion of PCA optimized by PSO and hedged GRU. The method aims to solve the problems in the existing technology that it is easily affected by threshold setting, the model extraction features are incomplete and inaccurate, the model is single, and the warning effect is poor.

[0009] In order to achieve the above technical effects, the technical solution adopted by the present invention is:

[0010] The temperature warning method for the slip ring chamber of a hydropower station based on the fusion of PCA and hedging GRU optimized by PSO includes the following steps:

[0011] S1, Data Collection:

[0012] Collect the temperature data of the slip ring chamber, including the slip ring surface temperature, infrared radiation and oil film thickness data, the slip ring system working status data, the internal temperature data and air density data of the slip ring chamber, the ambient temperature and humidity data, and the generator current, voltage and power data;

[0013] S2, data preprocessing:

[0014] The variance elimination method deletes data that is not relevant to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance, which means that the data does not change with temperature; then the data is normalized;

[0015] S3, data dimensionality reduction:

[0016] Principal component analysis (PCA) is used to reduce the dimensionality of the data, and the particle swarm algorithm is used to optimize the PCA algorithm. The weighted value of the retained information output by the PCA algorithm and the reconstruction error is used as the target, and α is the weight; the formula is:

[0017] (1-α)×retained information amount+α×reconstruction error;

[0018] The particle swarm algorithm is used to optimize the target and find the optimal dimension reduction. The training set and test set are divided into 8:2 ratios.

[0019] S4, building a temperature prediction model based on hedged GRU;

[0020] S5, input normal data into the temperature prediction model for training and perform temperature warning.

[0021] Preferably, in step S4, constructing a temperature prediction model based on the hedged GRU includes:

[0022] The reduced-dimensional data is used as the input of the attention mechanism. The softmax function is used to calculate the weights of each parameter after PCA dimensionality reduction, and then multiplied by the reduced-dimensional parameters to adjust the weights of each input parameter.

[0023] The weighted parameters will be used as the input of the gated recurrent unit GRU, which extracts the timing features;

[0024] Add average pooling layer and maximum pooling layer to GRU to extract differential features;

[0025] The extracted differential features are hedged and the same features are deleted to obtain discretized differential features;

[0026] Use the upsampling layer to expand the features to the specified size and weight them with the features obtained by the attention mechanism, further adjust the feature weights, and further weight them through the residual structure;

[0027] Introducing bidirectional GRU to learn differentiated temporal features, using weighted data as the input of bidirectional GRU, and extracting both forward and reverse temporal features at the same time;

[0028] Introducing the batch normalization layer to standardize feature distribution and accelerate convergence;

[0029] The temperature data is output through the fully connected layer.

[0030] Furthermore, the model introduces attention mechanism, pooling layer, hedging layer, and batch normalization layer to improve the model results, solving the shortcomings of the single model. At the same time, it can extract differentiated time series features, better characterize temperature characteristics, and help improve the prediction accuracy of the model.

[0031] Preferably, the gated recurrent unit formula is as follows:

[0032]

[0033] ht=GRU(ht-1,xt)

[0034] Where x t is the input of time step t, h t-1 is the hidden state at the previous moment, h t is the candidate hidden layer state, which can be understood as the new information at the current moment, W z 、W r , W is the corresponding weight matrix, ⊙ operation represents multiplication; where σ represents the sigmoid function, Z t represents the update signal, r t Represents the reset signal, tanh represents the tanh function, h t-1 and x t is the input of GRU, h t is the output of GRU.

[0035] Preferably, the bidirectional gated recurrent unit formula is as follows:

[0036]

[0037] Where x t is the input of time step t, h t-1 is the hidden state at the previous moment, h t is the candidate hidden layer state; where, is the output of the forward GRU, is the output of the reverse GRU, out t is the output of the bidirectional GRU.

[0038] Preferably, the average pooling layer formula is as follows:

[0039]

[0040] The formula for the maximum pooling layer is as follows:

[0041]

[0042] Where, is the value of the t-th neuron in the i-th feature matrix of the l+1-th layer, W is the width of the pooling area, and They are the output after maximum pooling and average pooling of the l+1th layer of neurons, avg means taking the average value of all features in the pooling area, and max means taking the maximum value of all features in the pooling area.

[0043] Preferably, the hedge layer formula is as follows:

[0044]

[0045] Where, The output after hedging calculation of the l+2th layer neurons.

[0046] Furthermore, the batch normalization layer formula is as follows:

[0047] Input: Values ​​of x over a mini-batch: β = {X1...m}; Parameters to be learned: γ, β;

[0048] Output: {yi=BNγ,β(xi)};

[0049]

[0050] Where m is the size of the mini-batch data, x i is the input data, u B is the mean of each feature, is the variance of each eigenvalue, ∈ is a small constant used to prevent division by zero, is the standardized data, y i is the data after linear transformation, γ is the scale factor, and β is the offset factor.

[0051] Preferably, in step S5, normal data is used to train the model constructed in step S4, a normal temperature prediction model is established, the residual of the predicted value under normal state and the predicted value under alarm state is calculated, and the maximum value of the residual is set as the fault alarm threshold under normal operating conditions of the slip ring chamber. An alarm is triggered when the residual exceeds the threshold.

[0052] The beneficial effects of the present invention are as follows:

[0053] 1. Traditional PCA requires manual search for dimensionality reduction, and the evaluation index is only one of the amount of information retained and the reconstruction error. The present invention uses the particle swarm algorithm to improve PCA, and uses the particle swarm algorithm to find the optimal dimension with the weighted values ​​of the two evaluation indicators as the target. Data dimensionality reduction can retain more data information while reducing the influence of irrelevant parameters on the prediction results. At the same time, combining the particle swarm algorithm with the PCA dimensionality reduction algorithm can improve the dimensionality reduction effect of PCA and eliminate the influence of manual settings.

[0054] 2. The present invention uses pooling layers and hedging layers to obtain discretized difference features, so that the data features learned by the model are more complete. The attention mechanism is used to focus the model's attention on more important parameters. The upsampling layer is used to make the output features reach the specified feature size to facilitate further weighting. The residual structure is used to avoid the loss of early data features, and at the same time enable the model to have a deeper network structure. A deeper network structure means that the model can learn deeper data features. The batch normalization layer is used to standardize the feature distribution and accelerate model convergence; and the residual threshold method used in the present invention is more flexible than the fixed threshold method. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of a flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] Example 1:

[0057] The temperature warning method for the slip ring chamber of a hydropower station based on the fusion of PCA and hedging GRU optimized by PSO includes the following steps:

[0058] S1, Data Collection:

[0059] Collect the temperature data of the slip ring chamber, including the slip ring surface temperature, infrared radiation and oil film thickness data, the slip ring system working status data, the internal temperature data and air density data of the slip ring chamber, the ambient temperature and humidity data, and the generator current, voltage and power data;

[0060] S2, data preprocessing:

[0061] The variance elimination method deletes data that is not relevant to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance, which means that the data does not change with temperature; then the data is normalized;

[0062] S3, data dimensionality reduction:

[0063] Principal component analysis (PCA) is used to reduce the dimensionality of the data, and the particle swarm algorithm is used to optimize the PCA algorithm. The weighted value of the retained information output by the PCA algorithm and the reconstruction error is used as the target, and α is the weight; the formula is:

[0064] (1-α)×retained information amount+α×reconstruction error;

[0065] The particle swarm algorithm is used to optimize the target and find the optimal dimension reduction. The training set and test set are divided into 8:2 ratios.

[0066] S4, building a temperature prediction model based on hedged GRU;

[0067] S5, input normal data into the temperature prediction model for training and perform temperature warning.

[0068] Preferably, in step S4, constructing a temperature prediction model based on the hedged GRU includes:

[0069] The reduced-dimensional data is used as the input of the attention mechanism. The softmax function is used to calculate the weights of each parameter after PCA dimensionality reduction, and then multiplied by the reduced-dimensional parameters to adjust the weights of each input parameter.

[0070] The weighted parameters will be used as the input of the gated recurrent unit GRU, which extracts the timing features;

[0071] Add average pooling layer and maximum pooling layer to GRU to extract differential features;

[0072] The extracted differential features are hedged and the same features are deleted to obtain discretized differential features;

[0073] Use the upsampling layer to expand the features to the specified size and weight them with the features obtained by the attention mechanism, further adjust the feature weights, and further weight them through the residual structure;

[0074] Introducing bidirectional GRU to learn differentiated temporal features, using weighted data as the input of bidirectional GRU, and extracting both forward and reverse temporal features at the same time;

[0075] Introducing the batch normalization layer to standardize feature distribution and accelerate convergence;

[0076] The temperature data is output through the fully connected layer.

[0077] Furthermore, the model introduces attention mechanism, pooling layer, hedging layer, and batch normalization layer to improve the model results, solving the shortcomings of the single model. At the same time, it can extract differentiated time series features, better characterize temperature characteristics, and help improve the prediction accuracy of the model.

[0078] Preferably, the gated recurrent unit formula is as follows:

[0079]

[0080] ht=GRU(ht-1,xt)

[0081] Where x t is the input of time step t, h t-1 is the hidden state at the previous moment, h t is the candidate hidden layer state, which can be understood as the new information at the current moment, W z 、W r , W is the corresponding weight matrix, ⊙ operation represents multiplication; where σ represents the sigmoid function, Z t represents the update signal, r t Represents the reset signal, tanh represents the tanh function, h t-1 and x t is the input of GRU, h t is the output of GRU.

[0082] Preferably, the bidirectional gated recurrent unit formula is as follows:

[0083]

[0084] Where x t is the input of time step t, h t-1 is the hidden state at the previous moment, h t is the candidate hidden layer state; where, is the output of the forward GRU, is the output of the reverse GRU, out t is the output of the bidirectional GRU.

[0085] Preferably, the average pooling layer formula is as follows:

[0086]

[0087] The formula for the maximum pooling layer is as follows:

[0088]

[0089] Where, is the value of the t-th neuron in the i-th feature matrix of the l+1-th layer, W is the width of the pooling area, and They are the output after maximum pooling and average pooling of the l+1th layer of neurons, avg means taking the average value of all features in the pooling area, and max means taking the maximum value of all features in the pooling area.

[0090] Preferably, the hedge layer formula is as follows:

[0091]

[0092] Where, The output after hedging calculation of the l+2th layer neurons.

[0093] Furthermore, the batch normalization layer formula is as follows:

[0094] Input: Values ​​of x over a mini-batch: β = {X1...m}; Parameters to be learned: γ, β;

[0095] Output: {yi=BNγ,β(xi)};

[0096]

[0097]

[0098] Where m is the size of the mini-batch data, x i is the input data, u B is the mean of each feature, is the variance of each eigenvalue, ∈ is a small constant used to prevent division by zero, is the standardized data, y i is the data after linear transformation, γ is the scale factor, and β is the offset factor.

[0099] Preferably, in step S5, normal data is used to train the model constructed in step S4, a normal temperature prediction model is established, the residual of the predicted value under normal state and the predicted value under alarm state is calculated, and the maximum value of the residual is set as the fault alarm threshold under normal operating conditions of the slip ring chamber. An alarm is triggered when the residual exceeds the threshold.

[0100] Example 2:

[0101] like Figure 1 As shown, this embodiment provides a specific implementation method of a hydropower station slip ring chamber temperature early warning method based on the fusion of PSO-optimized PCA and hedging GRU:

[0102] 1. Collect data related to the temperature of the slip ring chamber of the hydropower station;

[0103] 2. Use variance elimination method to delete data with variance 0;

[0104] 3. Data normalization;

[0105] 4. Use PCA to reduce the dimensionality of the data;

[0106] 5. Use the particle swarm optimization algorithm to find the optimal dimension reduction by taking the weighted value of the retained information output by PCA and the reconstruction error as the target;

[0107] 6. Use the softmax function to process the reduced-dimensional data, perform weighting, and adjust the weights of each parameter;

[0108] 7. Use the weighted data as the input of GRU to extract time series features;

[0109] 8. Introduce average pooling layer and maximum pooling layer to extract differential features;

[0110] 9. Introduce the hedge layer to extract discretized differentiation features;

[0111] 10. Introduce the upsampling layer to obtain the specified feature size for further weighting;

[0112] 11, further weighted using residual structure;

[0113] 12. The weighted data is used as the input of the bidirectional GRU to extract both forward and reverse time series features.

[0114] 13. Introduce batch normalization layer to standardize the distribution of data features;

[0115] 14. Use the fully connected layer to continue extracting data features;

[0116] 15. Output temperature data;

[0117] 16. Calculate the residual between the predicted value under normal state and the predicted value under alarm state, and take the maximum residual value as the threshold;

[0118] 17. If the residual exceeds the threshold, an alarm will be issued.

Claims

1. A hydropower station slip ring chamber temperature warning method based on the fusion of PCA optimized by PSO and hedging GRU is characterized by: The following steps are involved: S1, Data Collection: Collect the temperature data of the slip ring chamber, including the slip ring surface temperature, infrared radiation and oil film thickness data, the slip ring system working status data, the internal temperature data and air density data of the slip ring chamber, the ambient temperature and humidity data, and the generator current, voltage and power data; S2, data preprocessing: The variance elimination method deletes data that is not relevant to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance, which means that the data does not change with temperature; then the data is normalized; S3, data dimensionality reduction: Principal component analysis (PCA) is used to reduce the dimensionality of the data, and the particle swarm algorithm is used to optimize the PCA algorithm. The weighted value of the retained information output by the PCA algorithm and the reconstruction error is used as the target, and α is the weight. The formula is: (1-α) × amount of retained information + α × reconstruction error; Use the particle swarm algorithm to optimize the target and find the optimal dimension reduction; Divide the training set and test set into a fixed ratio; S4, builds a temperature prediction model based on the hedge GRU, including: The reduced-dimensional data is used as the input of the attention mechanism. The softmax function is used to calculate the weights of each parameter after PCA dimensionality reduction, and then multiplied by the reduced-dimensional parameters to adjust the weights of each input parameter. The weighted parameters will be used as the input of the gated recurrent unit GRU, which extracts the timing features; Add average pooling layer and maximum pooling layer to GRU to extract differential features; The extracted differential features are hedged and the same features are deleted to obtain discretized differential features; Use the upsampling layer to expand the features to the specified size and weight them with the features obtained by the attention mechanism, further adjust the feature weights, and further weight them through the residual structure; Introducing bidirectional GRU to learn differentiated temporal features, using weighted data as the input of bidirectional GRU, and extracting both forward and reverse temporal features at the same time; Introducing the batch normalization layer to standardize feature distribution and accelerate convergence; Output temperature data through the fully connected layer; S5, input normal data into the temperature prediction model for training and perform temperature warning.

2. The hydropower station slip ring chamber temperature early warning method based on the fusion of PCA and hedging GRU optimized by PSO according to claim 1 is characterized in that: The gated recurrent unit formula is as follows: ； Where, is the input of time step t, is the hidden state at the last moment, is the candidate hidden layer state, which can be understood as the new information at the current moment, W z 、W r , W is the corresponding weight matrix, and the ⊙ operation represents multiplication; where, σ express sigmoid function, Indicates the update signal, Represents the reset signal, tanh represents the tanh function, and is the input of GRU, is the output of GRU.

3. The hydropower station slip ring chamber temperature early warning method based on the fusion of PCA optimized by PSO and hedging GRU according to claim 2 is characterized in that: The formula for the bidirectional gated recurrent unit is as follows: ； Where, is the input of time step t, is the hidden state at the last moment, is the candidate hidden layer state; where, is the output of the forward GRU, is the output of the reverse GRU, is the output of the bidirectional GRU.

4. The hydropower station slip ring chamber temperature early warning method based on the fusion of PCA optimized by PSO and hedging GRU according to claim 3 is characterized in that: The average pooling layer formula is as follows: ； The formula for the maximum pooling layer is as follows: ； Where, For the l+1 Tier i The first feature matrix t The value of a neuron, W is the pooling area width, and Respectively l+1 The output of the layer neurons after maximum pooling calculation and the output after average pooling calculation, avg Indicates taking the average value of all features in the pooling area, max Indicates taking the maximum value of all features in the pooling area.

5. The hydropower station slip ring chamber temperature early warning method based on the fusion of PSO optimized PCA and hedging GRU according to claim 4 is characterized in that: The hedge layer formula is as follows: ； Where, For the l+2 The output of the layer neurons after hedging calculation.

6. The hydropower station slip ring chamber temperature early warning method based on the fusion of PSO optimized PCA and hedging GRU according to claim 1 is characterized in that: In step S5, the model constructed in step S4 is trained using normal data to establish a normal temperature prediction model. The residual between the predicted value under normal conditions and the predicted value under alarm conditions is calculated, and the maximum residual value is set as the fault alarm threshold under normal operating conditions of the slip ring chamber. An alarm is triggered when the residual exceeds the threshold.

Images