Method and system for detecting and analyzing operation state of power system

Abstract

The invention discloses an electric power system operation state detection analysis method and system. The method comprises the following steps: collecting time sequence data of an electric power system in real time; performing feature extraction and mining processing on the time series data by using an LSTM loop network to obtain a corresponding first feature vector or advanced representation of a data feature; inputting the advanced representation of the first feature vector or the data feature into a fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type; and adopting a reinforcement learning algorithm to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time running state of the fan. By adopting the embodiment of the invention, the problems of gradient explosion and disappearance when the sequence data is processed by the traditional neural network can be avoided, and the efficient detection and analysis of the operation state of the power system can be realized.

Description

Technical Field

[0001] The present invention relates to the field of power system operation status detection, and in particular to a power system operation status detection and analysis method and system. Background Art

[0002] Power system operation analysis is crucial in power system planning, operation and control, aiming to ensure stable, reliable and efficient operation of the power system. With the expansion of power scale and complexity of structure, traditional analysis methods face many challenges. At the same time, the development of intelligent and Internet of Things technologies provides an opportunity to improve the accuracy and efficiency of analysis for a large number of operation data.

[0003] At present, the main methods for power system operation analysis include Newton-Raphson method, PQ decomposition method, linearized power flow calculation, eigenvalue analysis method, PV curve, and least squares estimation method. Newton-Raphson method is a widely used power flow calculation method. It determines the system power flow distribution by iteratively solving a group of nonlinear equations. This method has a large amount of calculation, slow convergence speed, and is sensitive to initial values. PQ decomposition method simplifies the calculation by decomposing active power and reactive power. This method has low accuracy and is difficult to handle complex nonlinear characteristics. Linearized power flow calculation is applicable to high-voltage direct current transmission systems, which simplifies the nonlinear relationship of AC systems, but the influence of voltage and power has a limited scope of application. Eigenvalue analysis method establishes an evaluation benchmark value by calculating the Jacobian framework of the system. This method has high computational complexity and is difficult to apply in real time. PV curve, the model has a power-voltage relationship to analyze the voltage curve. This method is complex to calculate and difficult to handle large-scale system dynamic changes. Least squares estimation method uses real-time data and system models for state estimation. This method has high requirements for data accuracy and real-time performance and is sensitive to noise and data loss. In general, although the above methods have been widely used in power system operation analysis, they still have some shortcomings. For example, the model is highly dependent. Traditional methods rely on physical models and have poor adaptability to model parameters and system changes. The computational complexity is high. Large-scale power systems require huge amounts of computation and have poor real-time performance. Data utilization is insufficient and the large amount of operating data of smart grids cannot be fully utilized. The flexibility and robustness are poor. Traditional methods are difficult to adapt to the complexity and dynamic changes of the system. Summary of the invention

[0004] The embodiment of the present invention provides a method and system for detecting and analyzing the operating status of an electric power system, which avoids the gradient explosion and vanishing problems when a traditional neural network processes sequence data, and realizes efficient detection and analysis of the operating status of the electric power system.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for detecting and analyzing the operating status of a power system, comprising:

[0006] Real-time collection of time series data of power systems;

[0007] Using an LSTM recurrent network, feature extraction and mining are performed on the time series data to obtain a corresponding first feature vector or a high-level representation of data features;

[0008] Inputting the first feature vector or the high-level representation of the data feature into a fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type;

[0009] A reinforcement learning algorithm is adopted to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operating status of the fan.

[0010] The embodiment of the present invention is implemented to collect the time series data of the power system in real time, and then use the LSTM recurrent network to extract and mine the time series data collected in real time to obtain the corresponding first feature vector or the high-level representation of the data feature. The LSTM recurrent network can be used to deeply mine the long-term dependence and potential mode in the time series data, thereby avoiding the gradient explosion and disappearance problems when the traditional neural network processes the sequence data. Then, the first feature vector or the high-level representation of the data feature is input into the fault diagnosis model, so that the fault diagnosis model extracts and fuses the input data and classifies it and outputs the classification probability of the fan fault type. By applying the multi-scale attention fusion mechanism, extracting multi-scale features, and combining the adaptive weighted fusion method, the accuracy and robustness of fault diagnosis can be improved, and efficient detection and analysis of the operation status of the power system can be realized. In addition, a deep learning mechanism is introduced, and the structure and parameters of the fault diagnosis model are adaptively adjusted based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operation status of the fan, so that the fault diagnosis model can adaptively adjust the strategy according to the real-time data feedback, continuously optimize the model performance, and further improve the accuracy of fault diagnosis.

[0011] As a preferred solution, the fault diagnosis model includes a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; the first feature vector or the high-level representation of the data feature is input into the fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type, specifically:

[0012] Inputting the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs multi-scale rebalanced samples;

[0013] Taking the multi-scale rebalanced samples as input of the multi-scale feature extraction module, performing multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and outputting feature information of multiple scales;

[0014] Taking the feature information of the multiple scales as input of the multi-scale attention fusion module, the multi-scale attention fusion module performs weighted fusion on the feature information of the multiple scales based on adaptive attention weights, and outputs a second feature vector;

[0015] The second feature vector is input to the classification module, so that the classification module performs classification processing on the second feature vector and outputs the classification probability of the fan fault type.

[0016] The preferred solution of the embodiment of the present invention is implemented by performing data enhancement processing on the input data through the data enhancement module to generate multi-scale rebalanced samples, which helps to improve the diversity and balance of the data set, thereby reducing the model's bias towards minority class samples during training and improving the model's ability to identify different fault types. In addition, a multi-scale feature extraction module is used to capture the feature information of samples at different scales, so that the model can better understand and represent complex data patterns, especially when dealing with problems such as fan failures that may involve multi-level features, and can extract more comprehensive information. Then, a multi-scale attention fusion module is used to perform weighted fusion of feature information at multiple scales based on adaptive attention weights, which can dynamically adjust the importance of features at different scales, making the model more flexible and accurate when dealing with tasks with different feature importances.

[0017] As a preferred solution, the method for detecting and analyzing the operating status of a power system further includes:

[0018] Performing maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature, and then inputting the first feature vector or the high-level representation of the data feature after the maximum and minimum normalization processing into a power prediction model, so that the power prediction model outputs a corresponding power prediction result;

[0019] The power prediction model is obtained by training a LSTM-based wind turbine model using a training data set, wherein the training data set includes several groups of sample data, and the sample data includes wind speed, three-phase voltage and current at the wind turbine grid connection point, and wind turbine power.

[0020] According to a preferred solution for implementing the embodiment of the present invention, before using the power prediction model to realize power prediction, the first eigenvector or the high-level representation of the data feature is first subjected to maximum and minimum normalization processing, which can place different data in the same range level and eliminate the dimensional differences between different features, thereby facilitating the neural network model to explore the relationship between the features, improving the accuracy of the output fitting, and avoiding the excessive influence of certain features on the model training due to their large dimensions.

[0021] As a preferred solution, the LSTM-based wind turbine model includes 2 LSTM layers, 2 Dropout layers and 1 Dense layer.

[0022] The preferred solution of the embodiment of the present invention uses two LSTM layers and two Dropout layers to increase the complexity of the model, so that the model can learn more abstract and complex features and patterns. At the same time, the Dropout layer can randomly discard a part of the neurons, effectively preventing the model from overfitting and enhancing the generalization ability of the model, while the Dense layer (fully connected layer) is used to integrate and compress these features into one output, which is suitable for the final prediction task.

[0023] As a preferred solution, the maximum and minimum normalization processing is performed on the first feature vector or the high-level representation of the data feature, specifically:

[0024]

[0025] In the formula, χ min Represents the minimum value of the sample data in the training data set; χ max represents the maximum value of the sample data in the training data set; χ represents the original data; χ * Represents the data after maximum and minimum normalization processing.

[0026] The preferred scheme of implementing the embodiment of the present invention utilizes the minimum and maximum values ​​of the sample data in the training data set to perform maximum and minimum normalization processing on the first eigenvector corresponding to the time series data or the high-level representation of the data feature, which can make the output results of the model easier to combine with practical problems and improve the interpretability of the model.

[0027] As a preferred solution, the LSTM recurrent network includes a plurality of LSTM cells;

[0028] The internal state of the LSTM cell is updated as follows:

[0029]

[0030] In the formula, represents the internal state of the i-th LSTM cell at time t; fi (t) represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; U i,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U o represents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell.

[0031] According to a preferred solution for implementing the embodiment of the present invention, the LSTM recurrent network includes a plurality of LSTM cells with linear self-loop characteristics. The common hidden units in the traditional recurrent network are replaced by the recurrent links between the LSTM cells. At the same time, the sigmoid function is used as the activation function of the external input gate unit and the output gate. The external input gate unit is updated in a manner similar to the forget gate, and the output gate decides whether to release it according to the current situation, thereby realizing a nonlinear gating effect, thereby achieving a gradual accumulation and update effect when the LSTM is used to process time series data.

[0032] In order to solve the same technical problem, an embodiment of the present invention further provides a power system operation status detection and analysis system, comprising:

[0033] Data acquisition module, used to collect time series data of power system in real time;

[0034] A data preprocessing module, used to perform feature extraction and mining processing on the time series data using an LSTM recurrent network to obtain a corresponding first feature vector or a high-level representation of data features;

[0035] A fault diagnosis module, used for inputting the first feature vector or the high-level representation of the data feature into a fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type;

[0036] The adaptive adjustment module is used to adopt a reinforcement learning algorithm to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operating status of the fan.

[0037] As a preferred solution, the fault diagnosis model includes a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; the fault diagnosis module specifically includes a data enhancement unit, a feature extraction unit, a feature fusion unit and a classification processing unit;

[0038] The data enhancement unit is used to input the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs multi-scale rebalanced samples;

[0039] The feature extraction unit is used to use the multi-scale rebalanced samples as input of the multi-scale feature extraction module, perform multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and output feature information of multiple scales;

[0040] The feature fusion unit is used to use the feature information of the multiple scales as input of the multi-scale attention fusion module, perform weighted fusion on the feature information of the multiple scales based on the adaptive attention weight through the multi-scale attention fusion module, and output a second feature vector;

[0041] The classification processing unit is used to input the second feature vector into the classification module, so that the classification module performs classification processing on the second feature vector and outputs the classification probability of the fan fault type.

[0042] As a preferred solution, the power system operation status detection and analysis system further includes:

[0043] A power prediction module is used to perform maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature, and then input the first feature vector or the high-level representation of the data feature after the maximum and minimum normalization processing into a power prediction model, so that the power prediction model outputs a corresponding power prediction result; wherein, the power prediction model is obtained by training an LSTM-based wind turbine model using a training data set, and the training data set includes several groups of sample data, and the sample data includes wind speed, three-phase voltage and current at the wind turbine grid connection point, and wind turbine power.

[0044] As a preferred solution, the LSTM recurrent network includes a plurality of LSTM cells;

[0045] The internal state of the LSTM cell is updated as follows:

[0046]

[0047]

[0048] In the formula, represents the internal state of the i-th LSTM cell at time t; f i (t) represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; U i,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U orepresents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 : A flow chart of a method for detecting and analyzing the operating status of a power system provided in Embodiment 1 of the present invention;

[0050] Figure 2 : A structural diagram of a power system operation status detection and analysis system provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Embodiment one:

[0053] Please refer to Figure 1 , is a method for detecting and analyzing the operating status of a power system provided by an embodiment of the present invention, the method comprises steps S1 to S4, each step is specifically as follows:

[0054] Step S1, collecting time series data of the power system in real time.

[0055] Step S2, using the LSTM recurrent network to perform feature extraction and mining processing on the time series data to obtain the corresponding first feature vector or a high-level representation of the data feature.

[0056] It should be noted that the LSTM recurrent network is used to deeply mine the long-term dependencies and potential patterns in the time series data, thereby generating a series of feature vectors or high-level representations rich in key information. These data forms "processed" by the LSTM recurrent network not only retain the essence of the original time series data, but also eliminate redundancy and noise, making the time series data an indispensable input material for fault diagnosis models and power prediction models.

[0057] In the power system, data generally presents time series characteristics. Traditional neural networks are often limited to the information at the current time point when processing such data, and it is difficult to capture the full picture of the time series. The LSTM recurrent network not only inherits the external loop structure of the RNN, but also embeds a unique "LSTM unit" loop mechanism inside it. This design goes beyond the simple affine transformation and nonlinear activation combination. Similar to conventional recurrent networks, each unit of the LSTM has a unified input and output interface, but its core lies in the introduction of a more complex parameterized gating system, which finely controls the inflow and outflow of information, ensuring the effective accumulation and timely forgetting of information within the unit, thereby showing higher efficiency and stability when processing long sequence data.

[0058] As a preferred solution, the LSTM recurrent network includes a plurality of LSTM cells; wherein the internal state of the LSTM cells is updated in accordance with formula (1)(2)(3)(4). The most important component of the LSTM recurrent network is the state unit Construct a linear self-loop, see equation (2), the weight (or associated time constant) is determined by the forget gate f i (t) Control (time t and LSTM cell i), the sigmoid function sets the weight to a value between 0 and 1. LSTM cells replace ordinary hidden units in traditional recurrent networks through mutual cyclic connections, introducing a more complex processing mechanism. In this architecture, the input features are first calculated by conventional artificial neurons, and then, if they are evaluated as valid by the sigmoid input gate, their values ​​are allowed to be accumulated into the cell state. The cell state itself has the characteristics of linear self-loop, but its weight is not fixed, but dynamically controlled by another key component-the forget gate. In addition, the output of the LSTM cell is not released unconditionally, but the output gate decides whether to "release" it according to the current situation. It is worth noting that whether it is the input gate, the forget gate or the output gate, the sigmoid function is used as the activation function to achieve a nonlinear gating effect. The external input gate unit is updated in a similar way to the forget gate, and its own parameters See formula (3).

[0059]

[0060] In the formula, represents the internal state of the i-th LSTM cell at time t; f i (t) represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; Ui,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U o represents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell.

[0061] Step S3, inputting the first feature vector or the high-level representation of the data feature into the fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type.

[0062] With the rapid development of science and technology, research in the field of fault diagnosis has entered a new stage and has achieved many remarkable achievements. However, in the sub-field of wind turbine fault diagnosis, the complex and changeable operating environment has put forward higher requirements for diagnostic technology. In view of the fact that wind turbines are exposed to extreme conditions for a long time, the single-scale analysis method is stretched when dealing with environmental noise and it is difficult to ensure the robustness of diagnosis. Therefore, this embodiment introduces a multi-scale analysis framework, which aims to deeply analyze the multi-dimensional and multi-level characteristics of CMS vibration data, and through refined processing, dig out richer multi-scale information. On this basis, combined with advanced attention mechanisms, intelligent fusion and precise optimization of cross-scale features are realized, aiming to greatly improve the feature recognition ability, diagnostic accuracy and environmental adaptability of the fault diagnosis model, thereby enhancing the stability and reliability of the entire system. On the other hand, the performance of data-driven fault diagnosis models is deeply affected by the quality of the data set, especially the balance of the data. In practical applications, the scarcity of fault samples has become a major bottleneck, resulting in serious imbalance in data distribution. This imbalanced state makes the model tend to favor the majority class (normal state) during training, and may ignore the identification of the minority class (fault state), thereby limiting the comprehensive coverage and accuracy of fault diagnosis. To meet this challenge, this embodiment proposes MACNN-BiLSTM, a new fault diagnosis method that integrates a multi-scale attention mechanism. Through a sophisticated framework design, this method not only demonstrates the advantages of high stability and fast convergence, but also significantly improves the robustness and diagnostic accuracy of the model in complex environments.

[0063] As a preferred solution, the fault diagnosis model includes a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; step S3 includes steps S31 to S34, and each step is specifically as follows:

[0064] Step S31: input the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs multi-scale rebalanced samples.

[0065] Step S32: taking the multi-scale rebalanced samples as input of the multi-scale feature extraction module, performing multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and outputting feature information of multiple scales.

[0066] It should be noted that the introduction of a multi-scale mechanism and the input of multi-scale rebalanced samples into the multi-scale feature extraction module for multi-scale feature extraction can cleverly transform a single time series signal into a multi-level expression, comprehensively display the signal characteristics at different time granularities, enhance the richness and depth of feature representation, and improve model performance.

[0067] As an example, for CMS vibration signals, a one-dimensional maximum pooling technique is used to accurately capture the maximum value in the sliding window, retain the instantaneous energy characteristics while weakening noise interference, achieve multi-time scale reconstruction in the time-frequency domain, and comprehensively extract global and local features, laying a solid foundation for the leap in model performance. During the training process, given a set of vibration signals X = {χ 1 ,… n ,… N}. Where N is the length of the original input data, χ n is the nth vibration value of the original input signal. The time series of the corresponding scale is obtained by one-dimensional maximum pooling with different step sizes, and the final data length is N / S. For the multi-scale calculation, please refer to formula (6).

[0068]

[0069] Where M s,j Indicates the multi-scale calculation results; X j ′ represents the jth value of the time series obtained at a specific scale (defined by the step size S) after the maximum pooling operation; S is the step size, which represents the number of samples that the window moves each time when the maximum pooling operation is performed on the original signal X.

[0070] Step S33, taking the feature information of multiple scales as the input of the multi-scale attention fusion module, the multi-scale attention fusion module performs weighted fusion on the feature information of multiple scales based on the adaptive attention weight, and outputs a second feature vector.

[0071] Step S34: input the second feature vector to the classification module, so that the classification module performs classification processing on the second feature vector and outputs the probability of label classification, that is, the classification probability of the fan fault type.

[0072] In this embodiment, the multi-scale feature extraction module is composed of three CNN-BiLSTM sub-networks of different scales, in which the convolution block is used to extract local features and the BiLSTM is used to extract long-term temporal dependency information; the classification module is composed of a fully connected layer and a Softmax layer.

[0073] Step S4, using a reinforcement learning algorithm, based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operating status of the fan, adaptively adjust the structure and parameters of the fault diagnosis model.

[0074] It should be noted that in the deep learning exploration of fan fault diagnosis, this embodiment cleverly combines the essence of CNN and BiLSTM in the deep learning exploration of fan fault diagnosis, and constructs an innovative architecture. This architecture first relies on the powerful local feature extraction capability of CNN to refine key short-time series features from the original fan data. Subsequently, the BiLSTM module, with its bidirectional perspective, deeply analyzes the long-term impact of these features in the time dimension and weaves local features into a global time series map. In order to enhance the adaptability and stability of the model, this patent also introduces the MS-PGAN data enhancement technology, which generates a multi-scale and balanced sample set through a multi-level progressive generation network. These samples not only expand the training data, but also incorporate cross-scale information dimensions, helping the model capture more delicate and complex features in fan faults. The diversified data generated by MS-PGAN is directly injected into the multi-scale feature processing process, and through three parallel CNN-BiLSTM sub-networks, each deeply mines features at different time granularities. Finally, the multi-scale attention fusion strategy came into play. It intelligently integrates feature information from each scale based on dynamically adjusted attention weights to form a highly condensed and comprehensive feature representation. This feature vector then enters the classification module, and through the carefully designed fully connected layer and Softmax layer, the classification probability of the wind turbine fault type is accurately calculated. This diagnostic model, which combines MS-PGAN data enhancement, CNN and Bi LSTM, and multi-scale attention fusion technology, not only significantly improves the accuracy and efficiency of fault identification, but also provides strong technical support and guarantee for the operation and maintenance management of the wind power industry. In short, this model achieves in-depth mining and accurate diagnosis of wind turbine fault characteristics by cleverly combining CNN's local feature extraction and BiLSTM's global timing analysis capabilities.

[0075] As a preferred solution, an embodiment of the present invention provides a method for detecting and analyzing the operation status of a power system, further comprising a power prediction process, the process comprising steps S5 to S6, each of which is specifically as follows:

[0076] Step S5, performing maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature.

[0077] As a preferred solution, the maximum and minimum normalization processing mentioned in step S5 is specifically referred to in formula (7).

[0078]

[0079] In the formula, χ min Represents the minimum value of the sample data in the training data set; χ max represents the maximum value of the sample data in the training data set; χ represents the original data; χ *Represents the data after maximum and minimum normalization processing.

[0080] It should be noted that normalization technology is like a bridge that spans the gap of dimensional differences between input data, allowing data as different as voltage, current, wind speed, etc. to compete fairly on the same platform and receive equal attention for convergence. By carefully setting the boundaries of normalization, it is not only possible to accelerate the convergence of the neural network to the optimal solution, but also to enhance its generalization ability, so that the model can remain robust and accurate when facing unknown data. Therefore, normalization is not only a link in data processing, but also a key strategy to improve the performance and adaptability of neural networks.

[0081] Step S6: input the first feature vector or the high-level representation of the data feature after the maximum and minimum normalization processing into the power prediction model, so that the power prediction model outputs the corresponding power prediction result.

[0082] The power prediction model is obtained by training the LSTM-based wind turbine model using the training data set. The training data set includes several groups of sample data, including wind speed Three-phase voltage at wind turbine grid connection point With current And fan power Where j is the node number of different wind turbines.

[0083] It should be noted that the sample data in the training data set can be obtained from the historical operation data of the power system, or can be obtained through offline time domain simulation by simulation software. Among them, the training set includes data of both steady-state operation and transient operation conditions to improve the generalization ability of the model under various conditions.

[0084] As a preferred solution, the LSTM-based wind turbine model includes 2 LSTM layers, 2 Dropout layers and 1 Dense layer.

[0085] It should be noted that the LSTM layer is used to capture dependencies in time series data, the Dropout layer is used to prevent overfitting, and the Dense layer (commonly known as the fully connected layer) is used to generate the final prediction or classification results. The alternating use of the LSTM layer and the Dropout layer helps the model reduce overfitting during training, improves the generalization ability of the model, and thus improves the prediction accuracy of the wind turbine output power. As an example, the Dropout layer parameter is selected as 0.2, the hidden unit of the LSTM layer is selected as 80, the number of iterations is 50 epochs, and the batch-size is 64.

[0086] In this embodiment, root mean square error (RMSE) and mean absolute error (MAE) are selected as evaluation indicators of the power prediction model. The calculation formula of RMSE is shown in formula (8), and the calculation formula of MAE is shown in formula (9).

[0087]

[0088] In the formula, y or χ represents the true value; y * Or * represents the predicted value; N represents the number of samples.

[0089] Please refer to Figure 2 , is a structural diagram of a power system operation status detection and analysis system provided by an embodiment of the present invention, the system includes a data acquisition module M1, a data preprocessing module M2, a fault diagnosis module M3 and an adaptive adjustment module M4, and the specific details of each module are as follows:

[0090] Data acquisition module M1, used for real-time acquisition of time series data of the power system;

[0091] The data preprocessing module M2 is used to extract and mine the time series data using the LSTM recurrent network to obtain the corresponding first feature vector or a high-level representation of the data feature;

[0092] A fault diagnosis module M3 is used to input the first feature vector or the high-level representation of the data feature into the fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type;

[0093] The adaptive adjustment module M4 is used to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operation status of the fan by using a reinforcement learning algorithm.

[0094] As a preferred solution, the fault diagnosis model includes a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; the fault diagnosis module M3 specifically includes a data enhancement unit 31, a feature extraction unit 32, a feature fusion unit 33 and a classification processing unit 34, and each unit is specifically as follows:

[0095] A data enhancement unit 31, configured to input the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs a multi-scale rebalanced sample;

[0096] A feature extraction unit 32 is used to use the multi-scale rebalanced samples as input to the multi-scale feature extraction module, perform multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and output feature information of multiple scales;

[0097] A feature fusion unit 33 is used to use the feature information of multiple scales as the input of the multi-scale attention fusion module, perform weighted fusion on the feature information of multiple scales based on the adaptive attention weight through the multi-scale attention fusion module, and output a second feature vector;

[0098] The classification processing unit 34 is used to input the second feature vector into the classification module, so that the classification module performs classification processing on the second feature vector and outputs the classification probability of the fan fault type.

[0099] As a preferred option, please refer to Figure 2 The embodiment of the present invention provides a power system operation status detection and analysis system, which also includes a power prediction module M5, which is specifically as follows:

[0100] The power prediction module M5 is used to perform maximum and minimum normalization processing on the first eigenvector or the high-level representation of the data feature, and then input the first eigenvector or the high-level representation of the data feature after the maximum and minimum normalization processing into the power prediction model, so that the power prediction model outputs the corresponding power prediction result; wherein, the power prediction model is obtained by training the LSTM-based wind turbine model using a training data set, and the training data set includes several groups of sample data, and the sample data include wind speed, three-phase voltage and current at the wind turbine grid connection point, and wind turbine power.

[0101] As a preferred solution, the LSTM recurrent network includes a number of LSTM cells;

[0102] Among them, the internal state of the LSTM cell is updated as follows:

[0103]

[0104]

[0105] In the formula, represents the internal state of the i-th LSTM cell at time t; f i (t) represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; U i,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U o represents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell.

[0106] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0107] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0108] The present invention provides a method and system for detecting and analyzing the operation status of a power system, which collects the time series data of the power system in real time, and then uses an LSTM recurrent network to extract and mine the time series data collected in real time, so as to obtain a corresponding first feature vector or a high-level representation of data features, and can use the LSTM recurrent network to deeply mine the long-term dependence and potential mode in the time series data, thereby avoiding the gradient explosion and disappearance problems when the traditional neural network processes sequence data, and then inputs the first feature vector or the high-level representation of the data features into a fault diagnosis model, so that the fault diagnosis model extracts and fuses the input data and classifies it, and outputs the classification probability of the fan fault type, and extracts multi-scale features by applying a multi-scale attention fusion mechanism, and combines an adaptive weighted fusion method to improve the accuracy and robustness of fault diagnosis. In addition, a deep learning mechanism is introduced, and the structure and parameters of the fault diagnosis model are adaptively adjusted based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operation status of the fan, so that the fault diagnosis model can adaptively adjust the strategy according to the real-time data feedback, continuously optimize the model performance, and further improve the accuracy of fault diagnosis.

[0109] Furthermore, the LSTM recurrent network includes several LSTM cells with linear self-loop characteristics. The recurrent links between the LSTM cells replace the ordinary hidden units in the traditional recurrent network. At the same time, the sigmoid function is used as the activation function of the external input gate unit and the output gate. The external input gate unit is updated in a manner similar to the forgetting gate, and the output gate decides whether to release it according to the current situation, thereby realizing a nonlinear gating effect, thereby achieving a gradual accumulation and update effect when using LSTM to process time series data.

[0110] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting and analyzing the operating status of a power system, characterized in that: include: Real-time collection of time series data of power systems; Using an LSTM recurrent network, feature extraction and mining are performed on the time series data to obtain a corresponding first feature vector or a high-level representation of data features; Inputting the first feature vector or the high-level representation of the data feature into a fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type; A reinforcement learning algorithm is adopted to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operating status of the fan.

2. A method for detecting and analyzing the operating status of a power system according to claim 1, characterized in that: The fault diagnosis model comprises a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; the first feature vector or the high-level representation of the data feature is input into the fault diagnosis model so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type, specifically: Inputting the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs multi-scale rebalanced samples; Taking the multi-scale rebalanced samples as input of the multi-scale feature extraction module, performing multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and outputting feature information of multiple scales; Taking the feature information of the multiple scales as input of the multi-scale attention fusion module, the multi-scale attention fusion module performs weighted fusion on the feature information of the multiple scales based on adaptive attention weights, and outputs a second feature vector; The second feature vector is input to the classification module, so that the classification module performs classification processing on the second feature vector and outputs the classification probability of the fan fault type.

3. A method for detecting and analyzing the operating status of a power system according to claim 1, characterized in that: Also includes: Performing maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature, and then inputting the first feature vector or the high-level representation of the data feature after the maximum and minimum normalization processing into a power prediction model, so that the power prediction model outputs a corresponding power prediction result; The power prediction model is obtained by training a LSTM-based wind turbine model using a training data set, wherein the training data set includes several groups of sample data, and the sample data includes wind speed, three-phase voltage and current at the wind turbine grid connection point, and wind turbine power.

4. A method for detecting and analyzing the operating status of a power system according to claim 3, characterized in that: The LSTM-based wind turbine model includes 2 LSTM layers, 2 Dropout layers and 1 Dense layer.

5. A method for detecting and analyzing the operating status of a power system according to claim 3, characterized in that: The performing of maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature is specifically: In the formula, χ min Represents the minimum value of the sample data in the training data set; χ max represents the maximum value of the sample data in the training data set; χ represents the original data; χ represents the data after maximum and minimum normalization.

6. A method for detecting and analyzing the operation status of a power system according to claim 1, characterized in that: The LSTM recurrent network includes a plurality of LSTM cells; The internal state of the LSTM cell is updated as follows: In the formula, represents the internal state of the i-th LSTM cell at time t; f i (t) Represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; U i,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U o represents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell.

7. A power system operation status detection and analysis system, characterized in that: include: Data acquisition module, used to collect time series data of power system in real time; A data preprocessing module, used to perform feature extraction and mining processing on the time series data using an LSTM recurrent network to obtain a corresponding first feature vector or a high-level representation of data features; A fault diagnosis module, used for inputting the first feature vector or the high-level representation of the data feature into a fault diagnosis model, so that the fault diagnosis model performs multi-scale feature extraction and fusion and classification processing on the input data and outputs the classification probability of the fan fault type; The adaptive adjustment module is used to adopt a reinforcement learning algorithm to adaptively adjust the structure and parameters of the fault diagnosis model based on the classification probability of the fan fault type output by the fault diagnosis model and the real-time operating status of the fan.

8. A power system operation status detection and analysis system as claimed in claim 7, characterized in that: The fault diagnosis model comprises a data enhancement module, a multi-scale feature extraction module, a multi-scale attention fusion module and a classification module; the fault diagnosis module specifically comprises a data enhancement unit, a feature extraction unit, a feature fusion unit and a classification processing unit; The data enhancement unit is used to input the first feature vector or the high-level representation of the data feature into the data enhancement module, so that the data enhancement module performs data enhancement processing on the input data and outputs multi-scale rebalanced samples; The feature extraction unit is used to use the multi-scale rebalanced samples as input of the multi-scale feature extraction module, perform multi-scale feature extraction on the multi-scale rebalanced samples through the multi-scale feature extraction module, and output feature information of multiple scales; The feature fusion unit is used to use the feature information of the multiple scales as input of the multi-scale attention fusion module, perform weighted fusion on the feature information of the multiple scales based on the adaptive attention weight through the multi-scale attention fusion module, and output a second feature vector; The classification processing unit is used to input the second feature vector into the classification module, so that the classification module performs classification processing on the second feature vector and outputs the classification probability of the fan fault type.

9. A power system operation status detection and analysis system as claimed in claim 7, characterized in that: Also includes: A power prediction module is used to perform maximum and minimum normalization processing on the first feature vector or the high-level representation of the data feature, and then input the first feature vector or the high-level representation of the data feature after the maximum and minimum normalization processing into a power prediction model, so that the power prediction model outputs a corresponding power prediction result; wherein, the power prediction model is obtained by training an LSTM-based wind turbine model using a training data set, and the training data set includes several groups of sample data, and the sample data includes wind speed, three-phase voltage and current at the wind turbine grid connection point, and wind turbine power.

10. A power system operation status detection and analysis system as claimed in claim 7, characterized in that: The LSTM recurrent network includes a plurality of LSTM cells; The internal state of the LSTM cell is updated as follows: In the formula, represents the internal state of the i-th LSTM cell at time t; f i (t) Represents the self-loop weight of the i-th LSTM cell; represents the parameters of the external input gate unit of the i-th LSTM cell; σ(x) represents the activation function; b i represents the bias of the i-th LSTM cell; U i,j Represents the input weight between the i-th LSTM cell and the j-th LSTM; represents the input vector of the jth LSTM cell at time t; W i,j Represents the recurrent weight between the i-th LSTM cell and the j-th LSTM; represents the output vector of the jth LSTM cell at time (t-1); b f represents the bias of the forget gate; U f Represents the input weight of the forget gate; W f Represents the recurrent weight of the forget gate; Represents the bias of the external input gate unit of the i-th LSTM cell; Represents the input weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; represents the recurrent weight between the external input gate unit of the i-th LSTM cell and the external input gate unit of the j-th LSTM cell; b o represents the bias of the output gate of the i-th LSTM cell; U o represents the input weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell; W o Represents the recurrent weight between the output gate of the i-th LSTM cell and the output gate of the j-th LSTM cell.

Images