A power grid risk prediction method and system based on machine learning

Abstract

This invention discloses a machine learning-based method and system for predicting power grid risks, relating to the field of power safety technology. By acquiring monitoring data of main equipment in the target power grid and preprocessing it to obtain feature vectors, these feature vectors are input into a pre-trained state recognition model to obtain the failure probability of each main equipment. Based on the failure probabilities, key nodes exceeding a preset threshold are selected, corresponding failure scenarios are constructed, and failure exit simulations and consequence assessments are performed to obtain power grid risk analysis results. This invention utilizes a machine learning model to predict the failure probability of main equipment and, based on the prediction results, conducts targeted scenario construction and risk analysis, reducing the computational load of traditional traversal, improving analysis efficiency, achieving dynamic integration of equipment health status and system risk, and enhancing the real-time performance, accuracy, and scientific nature of scheduling decisions in risk assessment.

Description

Technical Field

[0001] This invention relates to the field of power grid security technology, and specifically to a power grid risk prediction method and system based on machine learning. Background Technology

[0002] With the large-scale integration of new energy power generation and the continuous expansion of the power grid, the power system's operational structure is becoming increasingly complex, and power flow distribution exhibits significant time-varying and uncertain characteristics. Transformers, circuit breakers, and transmission lines, as key equipment in the power grid, directly affect the safe and stable operation of the grid. Under conditions of long-term high-load operation and frequent adjustments to operating modes, transformers are susceptible to the combined effects of mechanical vibration, electromagnetic shock, temperature fluctuations, and environmental factors. Hidden dangers such as internal insulation aging, winding loosening, and core abnormalities gradually accumulate, and fault modes become more concealed and progressive. Circuit breakers, transmission lines, and other main power equipment are also affected by a combination of factors, including insulation aging, abnormal temperature rise, and insufficient safety distances.

[0003] Current power grid risk analysis methods primarily rely on periodic statistical analysis and offline simulation, often depending on human experience or pre-defined rules to assess equipment status. This approach fails to comprehensively reflect the dynamic characteristics of changes under complex operating environments. When conducting N-1 or N-2 static security analyses, dispatching departments focus on checking security constraints at the topology level, typically based on fixed equipment availability assumptions. Equipment-level health status information is not effectively incorporated into risk modeling, resulting in a lack of linkage between equipment risks and system-level risks, and making it difficult to identify critical high-risk nodes in a timely manner. Furthermore, N-1 or N-2 verification uses a traversal approach to construct fault scenarios, with computational load increasing exponentially with the scale of the power grid, without optimizing the risk analysis model based on equipment failure probabilities. Summary of the Invention

[0004] The purpose of this invention is to solve the problem mentioned in the background art that the existing N-1 or N-2 calculation models for power grid risk analysis have difficulty in timely identification of key high-risk nodes, and to propose a power grid risk prediction method and system based on machine learning.

[0005] A first aspect of this invention provides a machine learning-based method for predicting power grid risks, the method comprising:

[0006] Acquire monitoring data of all main equipment in the target power grid;

[0007] The monitoring data of the target device is preprocessed to obtain a feature vector; the target device is any main device in the target power grid.

[0008] The feature vector is input into the state recognition model corresponding to the target device to obtain the failure probability of the target device.

[0009] Devices with a failure probability greater than a preset threshold are identified as critical nodes, resulting in a set of critical nodes.

[0010] Based on the set of key nodes, a fault scenario is constructed, a fault exit simulation is performed, and the consequences are assessed to obtain the power grid risk analysis results.

[0011] By implementing this technical solution, machine learning models are used to predict the failure probability of main equipment. Based on the prediction results, targeted scenario construction and risk analysis are carried out, reducing the amount of traditional traversal calculations, improving analysis efficiency, realizing the dynamic integration of equipment health status and system risk, and improving the real-time performance, accuracy, and scientific nature of scheduling decisions.

[0012] Optionally, the target device can be any transformer; the monitoring data includes at least vibration signals.

[0013] The preprocessing of the monitoring data of the target device to obtain the feature vector includes:

[0014] The original vibration signal is denoised to obtain the target signal;

[0015] Perform time-domain analysis on the target signal to extract time-domain features including at least peak value, root mean square value, and kurtosis index;

[0016] The target signal is subjected to frequency domain analysis to extract frequency domain features including at least the fundamental frequency amplitude and the spectral centroid.

[0017] The time-domain features and the frequency-domain features are combined to obtain a feature vector.

[0018] By implementing this technical solution, time-domain and frequency-domain features of vibration signals can be extracted, which can characterize the transformer's operating state from multiple dimensions, enhance the expressive power and discriminative power of feature vectors, provide high-quality input for subsequent fault probability prediction, and improve the model's recognition accuracy and robustness.

[0019] Optionally, the denoising process of the original vibration signal to obtain the target signal includes:

[0020] Adaptive noise-complete ensemble empirical mode decomposition is performed on the original vibration signal to obtain multiple modal components;

[0021] Multiple modal components are sorted from high to low frequency, and the mutual information of adjacent two modal components is calculated in turn to obtain a mutual information sequence.

[0022] Based on the mutation characteristics of the mutual information sequence, pure noise components are identified and eliminated to obtain the retained modal components.

[0023] For the retained modal components, the energy entropy of each modal component is calculated, and the average value is taken as the entropy threshold. Modal components with energy entropy exceeding the entropy threshold are identified as noisy modes, and modal components with energy entropy not exceeding the entropy threshold are identified as fault characteristic modes.

[0024] For the noisy modes, an improved wavelet thresholding method is used for denoising to obtain the denoised noisy modes;

[0025] The target signal is obtained by reconstructing the signal based on the fault characteristic modes and the noise-containing modes after noise reduction.

[0026] By implementing this technical solution, which combines adaptive noise complete set empirical mode decomposition and energy entropy classification, the precise identification and hierarchical processing of noise components in vibration signals are achieved. This not only eliminates pure noise components but also preserves fault characteristic modes, avoiding the problems of excessive smoothing or feature loss in traditional denoising methods, and significantly improving the quality of signal reconstruction.

[0027] Optionally, determining and eliminating pure noise components based on the mutation characteristics of the mutual information sequence includes:

[0028] Calculate the difference between adjacent elements in the mutual information sequence to obtain the difference sequence D;

[0029] Based on the maximum value D in the difference sequence max Calculate the mutation threshold T: ; 'a' is a scaling factor less than 1;

[0030] The first difference D in the difference sequence that is greater than the mutation threshold T is... K The corresponding index K serves as the point before mutation;

[0031] The first K modal components are determined to be pure noise components.

[0032] By implementing this technical solution, the differential method is used to identify the abrupt change point of the mutual information sequence, which can adaptively determine the boundary of the pure noise component, avoid the blindness of manually setting thresholds, improve the accuracy and robustness of noise removal, and is applicable to vibration signals of different types and intensities.

[0033] Optionally, the threshold function used in the improved wavelet thresholding method is:

[0034] ;

[0035] in, These are the detail coefficients after thresholding. It is the detail coefficient at the j-th scale. It is the corresponding wavelet threshold; sgn() is the sign function; exp() is an exponential function with the natural constant e as the base; b is a constant greater than 0.

[0036] A second aspect of this invention provides a power grid risk prediction system based on machine learning, the system comprising:

[0037] The data acquisition module is used to acquire monitoring data of all main equipment in the target power grid;

[0038] The preprocessing module is used to preprocess the monitoring data of the target device to obtain a feature vector; the target device is any main device in the target power grid.

[0039] The fault prediction module is used to input the feature vector into the state recognition model corresponding to the target device to obtain the fault probability of the target device;

[0040] The key point determination module is used to identify devices with a failure probability greater than a preset threshold as key nodes, thereby obtaining a set of key nodes.

[0041] The risk simulation module is used to construct fault scenarios based on the set of key nodes, simulate fault exit, assess the consequences, and obtain power grid risk analysis results.

[0042] Optionally, the target device can be any transformer; the monitoring data includes at least vibration signals; the preprocessing module includes:

[0043] The noise reduction module is used to denoise the original vibration signal to obtain the target signal;

[0044] The time-domain feature extraction module is used to perform time-domain analysis on the target signal and extract time-domain features including at least peak value, root mean square value, and kurtosis index.

[0045] The frequency domain feature extraction module is used to perform frequency domain analysis on the target signal and extract frequency domain features including at least the fundamental frequency amplitude and the spectral centroid.

[0046] The feature combination module is used to combine the time-domain features and the frequency-domain features to obtain a feature vector.

[0047] Optionally, the noise reduction module includes:

[0048] The mode decomposition module is used to perform adaptive noise-complete ensemble empirical mode decomposition on the original vibration signal to obtain multiple modal components;

[0049] The noise removal module is used to sort multiple modal components from high to low frequency, calculate the mutual information of adjacent two-order modal components in turn, and obtain a mutual information sequence; based on the abrupt change characteristics of the mutual information sequence, it determines and removes pure noise components to obtain the retained modal components.

[0050] The classification module is used to calculate the energy entropy of each mode component for the retained mode components, and take the average value as the entropy threshold. Mode components with energy entropy exceeding the entropy threshold are identified as noisy modes, and mode components with energy entropy not exceeding the entropy threshold are identified as fault characteristic modes.

[0051] The residual noise suppression module is used to perform noise reduction on noisy modes using an improved wavelet thresholding method to obtain the noise-reduced noisy modes.

[0052] The signal reconstruction module is used to reconstruct the signal based on the fault characteristic mode and the noise-containing mode after noise reduction, so as to obtain the target signal.

[0053] Optionally, the noise removal module includes:

[0054] The difference calculation module is used to calculate the difference between adjacent elements in the mutual information sequence to obtain the difference sequence D;

[0055] The threshold calculation module is used to calculate the threshold value based on the maximum value D in the difference sequence. max Calculate the mutation threshold T: ; 'a' is a scaling factor less than 1;

[0056] The mutation determination module is used to identify the first difference D in the difference sequence that is greater than the mutation threshold T. K The corresponding index K serves as the point before mutation;

[0057] The noise determination module is used to identify the first K modal components as pure noise components.

[0058] Optionally, the threshold function used in the improved wavelet thresholding method is:

[0059] ;

[0060] in, These are the detail coefficients after thresholding. It is the detail coefficient at the j-th scale. It is the corresponding wavelet threshold; sgn() is the sign function; exp() is an exponential function with the natural constant e as the base; b is a constant greater than 0. Attached Figure Description

[0061] Figure 1 A flowchart illustrating a machine learning-based power grid risk prediction method provided in an embodiment of the present invention;

[0062] Figure 2 A threshold function comparison chart provided for an embodiment of the present invention;

[0063] Figure 3 This is an architecture diagram of a power grid risk prediction system based on machine learning, provided for an embodiment of the present invention. Detailed Implementation

[0064] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0065] This invention provides a machine learning-based method for predicting power grid risks. See also... Figure 1 , Figure 1 A flowchart illustrating a machine learning-based power grid risk prediction method provided in this embodiment of the invention. The method includes the following steps:

[0066] S101, acquire monitoring data of all main equipment in the target power grid.

[0067] S102, preprocess the monitoring data of the target device to obtain the feature vector.

[0068] S103, input the feature vector into the state recognition model corresponding to the target device to obtain the failure probability of the target device.

[0069] S104. Select devices with a failure probability greater than a preset threshold as key nodes to obtain a set of key nodes.

[0070] S105: Construct fault scenarios based on the set of key nodes, simulate fault exit, and conduct consequence assessment to obtain power grid risk analysis results.

[0071] The target equipment is any main equipment in the target power grid.

[0072] This invention provides a machine learning-based power grid risk prediction method. By using a machine learning model to screen high-risk critical nodes, it constructs targeted fault scenarios and conducts exit simulations and consequence assessments, significantly reducing computational load and improving analysis efficiency. Simultaneously, it incorporates real-time equipment health status into system-level risk modeling, achieving dynamic linkage between equipment risk and overall power grid risk, thus enhancing the foresight and precision of risk assessment.

[0073] In one implementation, the state recognition model can be either a support vector machine or a neural network model. Training data uses historical fault data and normal data, labeled as normal / abnormal, and the training objective is to maximize classification accuracy. A preset threshold can be set by technical personnel based on experience. For example, based on the distribution of model output scores of historical fault samples in the test set before the fault occurred, the threshold can be set slightly below its low quantile (e.g., the 5th quantile) to provide a certain safety margin, avoid the influence of extreme noise, and ensure coverage of typical fault characteristics.

[0074] In one embodiment, taking a transformer as an example, the monitoring data includes at least vibration signals. Step S102 involves preprocessing the monitoring data of the target device to obtain a feature vector including:

[0075] S1021, the original vibration signal is denoised to obtain the target signal.

[0076] S1022 performs time-domain analysis on the target signal to obtain time-domain characteristics, including peak value, root mean square value, kurtosis index, peak-to-peak value, skewness, etc.

[0077] S1023 performs frequency domain analysis on the target signal to obtain frequency domain characteristics, including fundamental frequency amplitude, spectral centroid, mean square frequency, etc.

[0078] S1024 combines the time-domain features and frequency-domain features to obtain the feature vector.

[0079] Because transformer vibration signals are easily affected by environmental noise and electromagnetic interference during acquisition, directly analyzing the raw vibration signals may reduce the accuracy of feature extraction. Therefore, denoising processing helps improve signal quality and the reliability of subsequent state identification. Furthermore, by extracting the time-domain and frequency-domain features of the vibration signal, the transformer's operating state can be characterized from multiple dimensions, such as energy distribution, impact characteristics, and frequency structure. This allows the feature vector to contain richer state information, which is beneficial for improving the accuracy and stability of the state identification model in determining the probability of transformer faults.

[0080] In one implementation, step S1021: denoising the original vibration signal to obtain the target signal includes:

[0081] Step 1: Perform adaptive noise complete set empirical mode decomposition on the original vibration signal to obtain multiple modal components.

[0082] Step 2: Sort the multiple modal components in descending order of frequency {IMF1, IMF2, ..., IMF} n}, calculate the mutual information of adjacent two-order modal components in turn to obtain the mutual information sequence X: {X1, X2, ..., X}n-1}, where X i =MI(IMF i IMF i+1 MI() represents the function for calculating mutual information.

[0083] Step 3: Based on the mutation characteristics of the mutual information sequence, determine and remove pure noise components to obtain the retained modal components.

[0084] Step 4: For the retained modal components, calculate the energy entropy of each modal component and take its average value as the entropy threshold. Modal components with energy entropy exceeding the entropy threshold are identified as noisy modes, and modal components with energy entropy not exceeding the entropy threshold are identified as fault characteristic modes.

[0085] Step 5: For the noisy mode, the wavelet thresholding method is used for noise reduction to obtain the noise-reduced noisy mode.

[0086] Step 6: Reconstruct the signal based on the fault characteristic modes and the noise-containing modes after noise reduction to obtain the target signal.

[0087] The mutual information values ​​between the noise-dominated first few modal components and their adjacent modal components are usually very small because they are mainly random noise with no informational correlation. However, modal components containing fault features form strong nonlinear information coupling due to sharing fault feature information, resulting in a sudden and significant jump in mutual information values. This abrupt change in information allows us to eliminate pure noise modal components without any fault features. Then, for the remaining effective modal components with a small amount of noise, we use energy entropy threshold detection to distinguish between fault feature modal components and noisy modal components. Secondary denoising is performed only on noisy modal components, while fault feature modal components are directly retained. This avoids over-denoising caused by uniform denoising of all modal components, ensuring denoising accuracy and feature fidelity, and thus providing a more stable and reliable data foundation for subsequent feature extraction and fault probability assessment.

[0088] In one implementation, determining and eliminating pure noise components based on the mutation characteristics of the mutual information sequence includes:

[0089] Step 1: Calculate the difference between adjacent elements in the mutual information sequence X to obtain the difference sequence D: D = {D1, D2, ..., D...} n-2}. Where D i =X i+1 -X i .

[0090] Step 2, based on the maximum value D in the difference sequence max Calculate the mutation threshold T: Where 'a' is a scaling factor less than 1, with values ​​between [0.3, 0.5].

[0091] Step 3: Select the first difference D in the difference sequence that is greater than the mutation threshold T. K The corresponding index K is used as the point before mutation.

[0092] Step 4: Convert the first K modal components IMF1-IMF K It was determined to be a pure noise component.

[0093] This embodiment employs differential calculation to adaptively determine the threshold, avoiding the subjectivity and arbitrariness of directly setting a fixed threshold. Furthermore, this method requires only simple calculations, has high computational efficiency, and exhibits good adaptability to different types of vibration signals.

[0094] In one implementation, wavelet thresholding is used to denoise the noisy modes, resulting in the following denoised noisy modes:

[0095] Step one involves using wavelet transform to decompose each noisy mode into detail coefficients and approximation coefficients at multiple scales. Specifically, a db6 wavelet basis can be used, with a decomposition level of 4.

[0096] Step two involves applying a preset threshold function to the detail coefficients at multiple scales to obtain multiple target detail coefficients. Specifically, this invention proposes an improved threshold function:

[0097] ;

[0098] in, These are the detail coefficients after thresholding. is the detail coefficient at the j-th scale; sgn() is the sign function; exp() is an exponential function with the natural constant e as the base; b is a constant greater than 0, used to adjust the steepness of the decay curve, and can be set to 1; The wavelet threshold for the j-th layer can be a traditional universal threshold (VisuShrink).

[0099] ;

[0100] in, is the median of the detail coefficients at the j-th layer; N is the signal length.

[0101] Step 3: Perform inverse wavelet transform on the approximation coefficients and multiple target detail coefficients to obtain the denoised noisy mode.

[0102] In traditional wavelet thresholding denoising, while hard thresholding functions can preserve local signal details, their discontinuities at the threshold can lead to oscillations and pseudo-Gibbs effects in the reconstructed signal. Soft thresholding functions, while offering good continuity, can cause constant-amplitude contraction of wavelet coefficients above the threshold, resulting in an overly smooth signal, particularly prone to losing subtle impact characteristics related to transformer faults. See also... Figure 2 , Figure 2 This invention provides a comparison chart of threshold functions according to an embodiment of the invention. The improved threshold function proposed in this embodiment introduces an exponential decay term, in... The region achieves a smooth transition, overcoming the discontinuity of the hard threshold function at the threshold point and ensuring that the processed coefficients approach the original coefficients infinitely as the original coefficients increase, significantly reducing the constant deviation of the soft threshold function. This design can effectively suppress noise while preserving the high-frequency fault characteristics reflecting the mechanical state of the transformer to the greatest extent, thereby obtaining a noise-reduced mode with a higher signal-to-noise ratio and more realistic details.

[0103] In one embodiment, the main equipment can also be a circuit breaker, transmission line, etc. Correspondingly, the monitoring data can also include leakage current signals, harmonic signals, temperature rise signals, safety distance information, and other relevant data. These data are analyzed and classified to determine the operating status of the main equipment.

[0104] In one embodiment, a fault scenario is constructed based on a set of critical nodes, a fault exit simulation is performed, and a consequence assessment is conducted to obtain the following power grid risk analysis results:

[0105] Step 1: Based on the set of key nodes, construct single-node fault scenarios and multi-node combined fault scenarios respectively. For single-node scenarios, simulate the condition where the transformer is taken out of operation due to a fault; for multi-node scenarios, they can be combined according to the fault probability or electrical distance to construct N-2 or Nk fault scenarios.

[0106] Step two involves performing power flow reconfiguration calculations and fault exit simulations based on the target power grid's topology and operating parameters under the constructed fault scenario, updating the voltage at each node, line power flow, and load distribution. The power redistribution after transformer shutdown can be analyzed using AC or DC power flow calculation methods to determine if line overload, voltage exceeding limits, islanding operation, or cascading failures have occurred.

[0107] Step 3: Quantitatively assess the consequences of the fault to form risk analysis results, which are used to assist in operation scheduling and maintenance decisions, including but not limited to: load loss rate; line or equipment overload rate; voltage qualification rate; changes in power supply reliability indicators; and the scale of potential cascading faults.

[0108] This invention provides a power grid risk prediction system based on machine learning. See also... Figure 3 , Figure 3 This is an architecture diagram of a machine learning-based power grid risk prediction system provided in an embodiment of the present invention. The system includes:

[0109] The data acquisition module is used to acquire monitoring data of all main equipment in the target power grid.

[0110] The preprocessing module is used to preprocess the monitoring data of the target device to obtain feature vectors.

[0111] The fault prediction module is used to input the feature vector into the state recognition model corresponding to the target device to obtain the fault probability of the target device.

[0112] The key point determination module is used to identify devices with a failure probability greater than a preset threshold as key nodes, thereby obtaining a set of key nodes.

[0113] The risk simulation module is used to construct fault scenarios based on the set of key nodes, simulate fault exit, assess the consequences, and obtain power grid risk analysis results.

[0114] The power grid risk prediction system based on machine learning provided in this invention uses machine learning models to screen high-risk critical nodes, constructs targeted fault scenarios, and conducts exit simulations and consequence assessments, significantly reducing computational load and improving analysis efficiency. Simultaneously, it incorporates real-time equipment health status into system-level risk modeling, achieving dynamic linkage between equipment risk and overall power grid risk, thus enhancing the foresight and precision of risk assessment.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.

Claims

1. A machine learning-based method for predicting power grid risks, characterized in that, The method includes: Acquire monitoring data of all main equipment in the target power grid; The monitoring data of the target device is preprocessed to obtain a feature vector; the target device is any main device in the target power grid. The feature vector is input into the state recognition model corresponding to the target device to obtain the failure probability of the target device. Devices with a failure probability greater than a preset threshold are identified as critical nodes, resulting in a set of critical nodes. Based on the set of key nodes, a fault scenario is constructed, a fault exit simulation is performed, and the consequences are assessed to obtain the power grid risk analysis results.

2. The power grid risk prediction method based on machine learning according to claim 1, characterized in that, The target device is any transformer; the monitoring data includes at least vibration signals. The preprocessing of the monitoring data of the target device to obtain the feature vector includes: The original vibration signal is denoised to obtain the target signal; Perform time-domain analysis on the target signal to extract time-domain features including at least peak value, root mean square value, and kurtosis index; The target signal is subjected to frequency domain analysis to extract frequency domain features including at least the fundamental frequency amplitude and the spectral centroid. The time-domain features and the frequency-domain features are combined to obtain a feature vector.

3. The power grid risk prediction method based on machine learning according to claim 2, characterized in that, The denoising process of the original vibration signal to obtain the target signal includes: Adaptive noise-complete ensemble empirical mode decomposition is performed on the original vibration signal to obtain multiple modal components; Multiple modal components are sorted from high to low frequency, and the mutual information of adjacent two modal components is calculated in turn to obtain a mutual information sequence. Based on the mutation characteristics of the mutual information sequence, pure noise components are identified and eliminated to obtain the retained modal components. For the retained modal components, the energy entropy of each modal component is calculated, and the average value is taken as the entropy threshold. Modal components with energy entropy exceeding the entropy threshold are identified as noisy modes, and modal components with energy entropy not exceeding the entropy threshold are identified as fault characteristic modes. For the noisy modes, an improved wavelet thresholding method is used for denoising to obtain the denoised noisy modes; The target signal is obtained by reconstructing the signal based on the fault characteristic modes and the noise-containing modes after noise reduction.

4. The power grid risk prediction method based on machine learning according to claim 3, characterized in that, The step of determining and eliminating pure noise components based on the mutation characteristics of the mutual information sequence includes: Calculate the difference between adjacent elements in the mutual information sequence to obtain the difference sequence D; Based on the maximum value D in the difference sequence max Calculate the mutation threshold T: ; 'a' is a scaling factor less than 1; the first difference value D in the difference value sequence that is greater than the mutation threshold T K the corresponding index K as a point before mutation; The first K modal components are determined to be pure noise components.

5. The power grid risk prediction method based on machine learning according to claim 3, characterized in that, The improved wavelet thresholding method uses the following threshold function: ； in, These are the detail coefficients after thresholding. It is the detail coefficient at the j-th scale. It is the corresponding wavelet threshold; sgn() is the sign function; exp() is an exponential function with the natural constant e as the base; b is a constant greater than 0.

6. A power grid risk prediction system based on machine learning, characterized in that, The system includes: The data acquisition module is used to acquire monitoring data of all main equipment in the target power grid; The preprocessing module is used to preprocess the monitoring data of the target device to obtain a feature vector; the target device is any main device in the target power grid. The fault prediction module is used to input the feature vector into the state recognition model corresponding to the target device to obtain the fault probability of the target device; The key point determination module is used to identify devices with a failure probability greater than a preset threshold as key nodes, thereby obtaining a set of key nodes. The risk simulation module is used to construct fault scenarios based on the set of key nodes, simulate fault exit, assess the consequences, and obtain power grid risk analysis results.

7. A power grid risk prediction system based on machine learning according to claim 6, characterized in that, The target device is any transformer; the monitoring data includes at least vibration signals; the preprocessing module includes: The noise reduction module is used to denoise the original vibration signal to obtain the target signal; The time-domain feature extraction module is used to perform time-domain analysis on the target signal and extract time-domain features including at least peak value, root mean square value, and kurtosis index. The frequency domain feature extraction module is used to perform frequency domain analysis on the target signal and extract frequency domain features including at least the fundamental frequency amplitude and the spectral centroid. The feature combination module is used to combine the time-domain features and the frequency-domain features to obtain a feature vector.

8. A power grid risk prediction system based on machine learning according to claim 7, characterized in that, The noise reduction module includes: The mode decomposition module is used to perform adaptive noise-complete ensemble empirical mode decomposition on the original vibration signal to obtain multiple modal components; The noise removal module is used to sort multiple modal components from high to low frequency, calculate the mutual information of adjacent two-order modal components in turn, and obtain a mutual information sequence; based on the abrupt change characteristics of the mutual information sequence, it determines and removes pure noise components to obtain the retained modal components. The classification module is used to calculate the energy entropy of each mode component for the retained mode components, and take the average value as the entropy threshold. Mode components with energy entropy exceeding the entropy threshold are identified as noisy modes, and mode components with energy entropy not exceeding the entropy threshold are identified as fault characteristic modes. The residual noise suppression module is used to perform noise reduction on noisy modes using an improved wavelet thresholding method to obtain the noise-reduced noisy modes. The signal reconstruction module is used to reconstruct the signal based on the fault characteristic mode and the noise-containing mode after noise reduction, so as to obtain the target signal.

9. A power grid risk prediction system based on machine learning according to claim 8, characterized in that, The noise removal module includes: The difference calculation module is used to calculate the difference between adjacent elements in the mutual information sequence to obtain the difference sequence D; The threshold calculation module is used to calculate the threshold value based on the maximum value D in the difference sequence. max Calculate the mutation threshold T: ; 'a' is a scaling factor less than 1; The mutation determination module is used to identify the first difference D in the difference sequence that is greater than the mutation threshold T. K The corresponding index K serves as the point before mutation; The noise determination module is used to identify the first K modal components as pure noise components.

10. A power grid risk prediction system based on machine learning according to claim 8, characterized in that, The improved wavelet thresholding method uses the following threshold function: ； in, These are the detail coefficients after thresholding. It is the detail coefficient at the j-th scale. It is the corresponding wavelet threshold; sgn() is the sign function; exp() is an exponential function with the natural constant e as the base; b is a constant greater than 0.

Images