Intelligent fault detection method and system for power distribution ring main unit

Abstract

The present invention relates to the technical field of fault detection of power distribution ring main units. Disclosed are an intelligent fault detection method and system for a power distribution ring main unit. The method comprises: acquiring real-time operation data of a power distribution ring main unit, and on the basis of the fluctuation degree of the real-time operation data, determining whether to trigger fault detection logic; in response to triggering the fault detection logic, inputting the real-time operation data into a pre-established fault detection model to obtain a fault detection result, the fault detection model being constructed by a multi-modal fusion-based convolutional long short-term memory network and being obtained by performing training by using an improved adaptive gradient modulation algorithm; and generating an alert on the basis of the fault detection result. In the present invention, a multi-modal fusion-based convolutional long short-term memory network is designed, and the contribution difference of each modality is adaptively adjusted to optimize a model training process, thereby improving the prediction accuracy of a model, and further ensuring the operation safety and stability of a power distribution ring main unit.

Description

A method and system for intelligent fault detection in power distribution ring main units Technical Field

[0001] This invention relates to the field of fault detection technology for power distribution ring main units, and in particular to an intelligent fault detection method and system for power distribution ring main units. Background Technology

[0002] Modern society's demand for electricity supply is increasing daily, and traditional distribution networks, due to their reliance on manual inspection and maintenance and low levels of automation and intelligence, are gradually revealing their limitations. Ring main units, as a key component of ring distribution networks, construct power supply lines into a closed loop structure, allowing each branch to obtain power from two directions. Even if some power supply lines fail, the relevant distribution branches can still obtain power from the other side of the ring network. This design not only ensures the reliability and continuity of power supply, but also, theoretically, provides a single-circuit power supply, effectively functioning as a dual-circuit power supply.

[0003] While the design concept of ring main units (RMMs) has greatly improved the stability of power supply, existing systems mainly rely on basic remote monitoring technologies. Although current technologies have made some progress in remote monitoring and fault handling of distribution RMMs, several problems and limitations remain. Current systems primarily rely on basic remote monitoring technologies, which, while capable of status monitoring and fault location, lack sufficient intelligence and struggle to respond quickly and effectively to complex and ever-changing power grid conditions. For example, traditional statistical methods such as linear regression, exponential smoothing, and autoregressive moving averages, while simple in principle and easy to implement, have limited predictive effectiveness when dealing with complex data, failing to meet the predictive needs of dynamic power grid changes. Deep learning models, on the other hand, suffer from reduced prediction accuracy when input features contain a large amount of discontinuous discrete data, and relying on a single model for prediction can lead to poor generalization ability, making it difficult to adapt to diverse power grid conditions. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent fault detection method, system, device, and storage medium for power distribution ring main units. This addresses the issues of untimely fault handling response and low prediction accuracy in existing power distribution ring main units, thereby improving fault prediction accuracy and efficiency and enhancing the operational safety of the power distribution ring main units.

[0005] In a first aspect, the present invention provides an intelligent fault detection method for a power distribution ring main unit, the method comprising:

[0006] Obtain real-time operating data of the power distribution ring network cabinet, and determine whether to trigger fault detection logic based on the fluctuation level of the real-time operating data;

[0007] In response to triggering the fault detection logic, the real-time running data is input into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed from a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm.

[0008] Based on the fault detection results, an alarm notification is generated;

[0009] The training steps of the fault detection model include:

[0010] Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met.

[0011] If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio.

[0012] The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient.

[0013] Based on the modulation coefficients, the updated learning rate, and the updated model gradient, the model parameters corresponding to each modality data are updated, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met.

[0014] Furthermore, the fault detection model includes multiple attention-based convolutional long short-term memory networks and a multimodal fusion module, which includes a linear addition layer and an output layer;

[0015] The number of attention-based convolutional long short-term memory networks is consistent with the number of data types of the frequency domain data. The input data of each attention-based convolutional long short-term memory network is the corresponding type of frequency domain data. The input data of the multimodal fusion module is the output data of each attention-based convolutional long short-term memory network. The output data of the multimodal fusion module is the fault detection result.

[0016] Furthermore, the step of acquiring historical operating data of the power distribution ring network cabinet and processing the data to obtain frequency domain data includes:

[0017] Obtain historical operating data of the power distribution ring main unit, including voltage data, current data, and temperature data;

[0018] The historical operational data is preprocessed, including missing value imputation, outlier handling, data cleaning, and standardization.

[0019] The historical running data after data preprocessing is converted into corresponding frequency domain data using a time-frequency conversion algorithm, which includes a fast Fourier transform algorithm based on a mixed basis and a continuous wavelet transform algorithm.

[0020] Furthermore, the step of obtaining the modal data corresponding to each frequency domain data based on the output of the convolutional long short-term memory network includes:

[0021] Feature data corresponding to each frequency domain data is obtained from each of the attention-based convolutional long short-term memory networks, and modal weights of each frequency domain data are calculated according to the attention mechanism.

[0022] Based on the feature data and the mode weights, the modal data corresponding to each frequency domain data is obtained, and the modal data includes voltage mode data, current mode data and temperature mode data.

[0023] Furthermore, the step of calculating the modulation coefficients of each mode data based on the mode contribution ratio includes:

[0024] Calculate the modal contribution ratio corresponding to each modal data based on the proportional relationship between each modal data and the other modal data;

[0025] Based on the modal contribution ratio, the modulation coefficients corresponding to each modal data are calculated.

[0026] Furthermore, the voltage mode data is represented by the following formula:

[0027] The current mode data are represented by the following formula:

[0028] The temperature modal data are represented by the following formula:

[0029] In the formula, This represents the u-mode data obtained in the i-th iteration, where u∈{V,C,T}, and V, C, and T represent voltage, current, and temperature, respectively. This represents the u-mode weights at step t. Let b represent the weight matrix of the u-mode data at step t. u This represents the bias term of the u-mode data. θ represents the feature extraction function for data of type u. u This represents the input parameters of the feature extraction function for u-type data. This represents the frequency domain data of type u input during the i-th iteration, where i represents the iteration number and t represents the number of iteration steps.

[0030] Furthermore, the modal contribution ratio is expressed by the following formula:

[0031] In the formula, This represents the modal contribution ratio of the u-mode data at step t. This represents the u-mode data obtained in the i-th iteration, u∈{V,C,T}, where V, C, and T represent voltage, current, and temperature, respectively. t This represents the random mini-batch selected at step t, where i represents the iteration number and t represents the number of iteration steps.

[0032] The modulation coefficient is expressed by the following formula:

[0033] In the formula, α represents the modulation coefficient of the u-mode data, and α represents the hyperparameter controlling the degree of modulation.

[0034] The model parameters are represented by the following formula:

[0035] In the formula, η represents the model parameters of the model corresponding to the input data of type u at step t. t It represents the learning rate updated in the current step, where t is the current iteration number. This represents the updated model gradient.

[0036] Furthermore, the step of determining whether to trigger fault detection logic based on the fluctuation level of the real-time operating data includes:

[0037] Calculate the standard deviation of the real-time running data and determine whether the standard deviation is greater than a preset threshold. If so, trigger the fault detection logic.

[0038] Furthermore, after the step of generating an alarm notification based on the fault detection result, the method further includes:

[0039] An emergency plan library containing multiple emergency plans is pre-built, and the priority of each emergency plan is set according to the scope of the failure's impact and the urgency of the failure;

[0040] Based on the alarm level indicated by the alarm, the corresponding priority emergency plan is automatically selected and executed from the emergency plan library.

[0041] Secondly, the present invention provides an intelligent fault detection system for a power distribution ring main unit, the system comprising:

[0042] The triggering module is used to acquire real-time operating data of the power distribution ring network cabinet and determine whether to trigger the fault detection logic based on the fluctuation level of the real-time operating data.

[0043] The fault detection module is used to respond to the triggering of fault detection logic by inputting the real-time running data into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed by a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm.

[0044] The training steps for the fault detection model include:

[0045] Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met.

[0046] If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio.

[0047] The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient.

[0048] The model parameters corresponding to each modality data are updated based on the modulation coefficient, the updated learning rate, and the updated model gradient, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met.

[0049] The alarm notification module is used to generate alarm notifications based on the fault detection results.

[0050] This invention provides an intelligent fault detection method and system for distribution ring main units. By designing a convolutional long short-term memory network based on multimodal fusion and optimizing the model training process through adaptive adjustment of the contribution differences of each modality, the invention improves the model's prediction accuracy. Furthermore, the fault detection model provided by this invention has scalability; for the needs of multiple input modalities, by modifying the modal contribution differences accordingly, it can fuse information from multiple modalities for learning. This allows the model to be widely applied in different distribution ring main unit scenarios, exhibiting high flexibility and scalability, further ensuring the safety and stability of distribution ring main unit operation. Attached Figure Description

[0051] Figure 1 is a flowchart illustrating the intelligent fault detection method for power distribution ring main unit in an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of the power grid state change when the distribution ring network cabinet fails in an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of the fault detection model in an embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of the single-step prediction structure of the fault detection model in an embodiment of the present invention;

[0055] Figure 5 is a schematic diagram of the intelligent fault detection system of the power distribution ring network cabinet in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please refer to Figure 1. The first embodiment of the present invention proposes an intelligent fault detection method for a power distribution ring main unit, which includes steps S10 to S30:

[0058] Step S10: Obtain the real-time operating data of the power distribution ring network cabinet, and determine whether to trigger the fault detection logic based on the fluctuation level of the real-time operating data;

[0059] Step S20: In response to triggering the fault detection logic, the real-time running data is processed, and the processed real-time running data is input into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed by a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm.

[0060] Step S30: Generate an alarm notification based on the fault detection results.

[0061] This invention efficiently captures key performance indicators of a power distribution ring main unit (RNB) by installing sensors on the RNB. Specifically, the following intelligent sensing devices are deployed in the RNB unit: current transformers for measuring line current, voltage sensors for monitoring line voltage, temperature and humidity sensors for recording ambient temperature and humidity, and insulation monitoring sensors for assessing the insulation status of the RNB. The deployment of these sensing devices fully considers the complexity of the power network environment, including the influence of external factors such as electromagnetic interference, temperature changes, and mechanical vibration, ensuring that the devices have sufficient anti-interference capabilities.

[0062] The collected data can be represented as a vector X containing n data points: X = {x1, x2, ..., xn} n}

[0063] Where, x i This represents the i-th data.

[0064] Please refer to Figure 2. When a fault occurs in the distribution ring main unit, the electrical and switching quantities of the power grid, such as voltage, current, and circuit breakers, will change. In order to achieve timely response to the fault and reduce the computational load of fault detection, this embodiment continuously monitors the system performance of the distribution ring main unit through a real-time data collection and analysis mechanism, and uses an anomaly detection algorithm to detect fault signs in real time. Specifically, this embodiment uses an anomaly detection algorithm based on standard deviation to detect fault signs. The formula for calculating the standard deviation is as follows:

[0065] In the formula, σ represents the standard deviation, and x i This represents the i-th data. represents the average value of the data, and n represents the total number of data points.

[0066] Then, it is determined whether the standard deviation exceeds the threshold. If it does, it indicates that the fluctuation of the data exceeds the normal range, suggesting a potential fault. At this point, the fault detection logic can be triggered. This embodiment does not require real-time fault prediction of the data; instead, it triggers the fault detection logic through anomaly detection. This ensures timely response to fault occurrences while effectively reducing the system load.

[0067] When the fault detection logic is triggered, based on the real-time operating data of the power distribution ring main unit, this embodiment uses a fault detection model constructed based on a multimodal fusion convolutional long short-term memory network to predict the faults of the power distribution ring main unit. The architecture of the fault detection model and its training steps are described in detail below.

[0068] First, in this embodiment, the dataset used to train the model is constructed using the following steps:

[0069] Obtain historical operating data of the power distribution ring main unit, including voltage data, current data, and temperature data;

[0070] The historical operational data is preprocessed, including missing value imputation, outlier handling, data cleaning, and standardization.

[0071] The historical running data after data preprocessing is converted into corresponding frequency domain data using a time-frequency conversion algorithm, which includes a fast Fourier transform algorithm based on a mixed basis and a continuous wavelet transform algorithm.

[0072] In this embodiment, historical operating data of the distribution ring main unit, including voltage, current, and temperature data, is acquired. This historical operating data is first preprocessed, including missing value imputation, outlier handling, data cleaning, and standardization. Specifically, missing value imputation is performed using regression imputation, a method based on a regression model. It utilizes the relationship between existing data features and the target attribute to select an appropriate regression model to predict missing attribute values. For example, linear regression or logistic regression is used for linear relationships; multinomial regression, decision tree regression, random forest regression, or support vector regression are used for nonlinear relationships. For time series characteristics, time series analysis models, such as autoregressive moving average models, are used for imputation.

[0073] The data is standardized using Z-Score Normalization. For each data point X, the mean μ and standard deviation σ are calculated, and the data is then converted into a Z-score.

[0074] The resulting new dataset will have a mean of 0 and a standard deviation of 1, making data of different magnitudes or units comparable.

[0075] Outlier handling is performed using the IQR (Interquartile Range) method. The first quartile (Q1) and the third quartile (Q3) of the dataset are calculated, and then the difference between these two quartiles, i.e., the IQR, is calculated: IQR = Q3 - Q1

[0076] The upper and lower bounds used to identify outliers are calculated based on the IQR and defined using the following formula:

[0077] LowerBound: LB=Q1-1.5*IQR

[0078] Upper Bound: UB = Q1 + 1.5 * IQR

[0079] Any data point below the lower bound or above the upper bound can be considered an outlier. These are the data points that require further analysis and may need to be excluded or processed in other ways.

[0080] Furthermore, based on logical rules or model output, obviously erroneous data is corrected or deleted. For example, reasonable range rules are used to correct data and remove duplicate records from the dataset. Of course, other data preprocessing methods can also be used to correct real-time running data, which will not be elaborated here.

[0081] After preprocessing the historical operational data, a time-frequency conversion algorithm is used to convert the historical operational data from time-domain data to the corresponding frequency-domain data. In this embodiment, the time-frequency conversion algorithm can be either the Fast Fourier Transform (FFT) algorithm based on mixed bases or the continuous wavelet transform algorithm. Of course, other algorithms can also be used, and no specific limitation is made here. When using the Fast Fourier Transform algorithm based on mixed bases, its mixed-basis FFT formula of length N can be expressed as:

[0082] In the formula, X[] represents the frequency domain sequence, x[] represents the time domain sequence, and N = N1 × N2. This formula means that the FFT problem is split into two smaller Discrete Fourier Transform (DFT) problems, and DFTs of length N1 and N2 are calculated respectively. Then the results are combined, and k1 and k2 represent the frequency domain data corresponding to the time domain data n1 and n2. Finally, DFT transformation is performed on each column of the processed result to obtain the final frequency domain representation.

[0083] When using the continuous wavelet transform algorithm, the continuous wavelet transform performs local analysis of the signal at various scales and locations to observe the frequency changes of the signal in the time series. The specific transform steps are similar to those of the mixed-basis fast Fourier transform algorithm or the standard steps of the continuous wavelet transform algorithm, and will not be repeated here.

[0084] By using a time-frequency conversion algorithm, historical operating data is converted from the time domain to the frequency domain, resulting in three types of frequency domain data: voltage, current, and temperature. A dataset is then built based on these three types of frequency domain data, and the fault detection model with the above structure is trained using this dataset to obtain a trained fault detection model.

[0085] In this embodiment, the fault detection model includes multiple attention-based Convolutional Long Short-Term Memory (ConvLSTM) networks and a multimodal fusion module. Its model architecture is shown in Figure 3. The ConvLSTM networks consist of a Convolutional Neural Network (CNN) model and a Long Short-Term Memory (LSTM) model, introducing convolutional operations on top of the original LSTM. While traditional LSTM models process data through fully connected layers, ConvLSTM uses convolutional layers to process data, thus better capturing spatial features in the data. The number of ConvLSTM networks corresponds to the number of data types in the input data. The multimodal fusion module fuses the predicted values ​​output by each ConvLSTM network through a linear summation layer and then classifies the fused predicted values ​​through an output layer, thus outputting the fault detection result.

[0086] The reason for using a Convolutional Long Short-Term Memory (CNN) network to construct the fault detection model in this embodiment is that fault detection in distribution ring main units is a typical spatiotemporal sequence prediction problem, requiring simultaneous consideration of long-term temporal dependencies and local spatial correlations. The CNN model can fully utilize these characteristics of distribution network data, thereby achieving higher accuracy and efficiency in fault prediction. Specifically, the CNN model in this embodiment includes multiple convolutional and pooling layers to extract global features from the processed data. The convolutional layers are responsible for extracting spatial features from the data, identifying key characteristics in the signal by capturing local dependencies. The pooling layers downsample the output of the convolutional layers to reduce computation and improve the generalization ability of the features. The data after feature extraction by the CNN is input into an LSTM model for further analysis. The LSTM model, through its convolutionally gated recurrent units, can capture and maintain temporal dependencies at each time step, which is particularly effective for processing and predicting systems with complex temporal dynamics. Furthermore, compared to the conventional ConvLSTM model structure, the convolutional long short-term memory network in this embodiment adds an attention layer. This attention layer is composed of an attention mechanism and a fully connected layer. The attention layer is placed after the LSTM model. The output of the hidden layer of the LSTM model is fed into the attention layer for processing. The attention layer can enhance the role of important time steps in the LSTM, thereby further reducing the model's prediction error. The attention layer is essentially a weighted summation of the LSTM's output vector, which is then trained through a fully connected layer and finally outputs the fault detection result through the output layer.

[0087] The single-step prediction structure of the ConvLSTM model is shown in Figure 4, and its basic formula for single-step prediction can be expressed as:

[0088] Among them, input gate I t Determine the input X at the current moment t How much information needs to be transmitted to the cell state? Here, W with subscripts represents the corresponding weight matrix, and H... t-1 It is the hidden state from the previous moment, C t-1 It represents the cell state at the previous moment, b I This is the bias term of the input gate. The result is compressed to the [0,1] interval using the sigmoid activation function σ, representing how much new information is retained. * indicates the convolution operation. This indicates an element-wise multiplication operation.

[0089] Among them, the forgetting gate F t Controls how much information is discarded from the cell state. Similar to the input gate, the forget gate uses the sigmoid activation function, W... XF and W HFAs the weight matrix and input X t And the previous hidden state H t-1 Perform convolution processing, C t-1 It is the cell state that determines which past information is discarded.

[0090] Among them, cell state C t This is the core part of the LSTM, storing long-term memory. Here... Indicates retained past information, I t The proportion of new information added is determined by the hyperbolic tangent (tanh) activation function, which compresses the information to the range [-1, 1]. This formula updates the cell state, combining past states with current input information.

[0091] Among them, output gate O t It determines which information needs to be output at the current moment. Similar to the input gate and forget gate, the sigmoid function controls the proportion of output. t The output gate combines the current input and cell state to generate the output, representing the updated cell state. In the above formulas, 'b' represents the bias term of the corresponding formula, and 'W' represents the convolution operation of the corresponding formula.

[0092] Among them, the hidden state H t It is the final output of the LSTM and also the input to the next time step of the recurrent neural network. It passes through the output gate O. t and the updated cell state C t Generation. The complete prediction steps can be found by referring to the prediction steps of a regular ConvLSTM model, and will not be repeated here.

[0093] As described above, the dataset in this embodiment includes three types of frequency domain data. Each type of frequency domain data is used by a corresponding attention-based convolutional long short-term memory network for fault prediction, thereby obtaining the predicted value independently predicted by each data source. These predicted values ​​contribute to the fault prediction of the distribution ring main unit to different degrees. In order to fuse the prediction results corresponding to different modal data and further enhance the prediction capability, this invention adds a multimodal fusion module to the ConvLSTM network model. This module consists of a linear addition layer and an output layer, which is used to fuse the predicted values ​​independently predicted by each data source, thereby improving the accuracy of model prediction.

[0094] The fault prediction model based on multimodal fusion provided by this invention is not only reflected in the model architecture, but also in the design of a corresponding model training method. By improving the model's fusion effect on multimodal information, the model can make fuller use of this information when accepting a wide range of multimodal information, further enriching the model's information sources and providing more comprehensive data support for accurately judging the fault status of distribution ring network cabinets.

[0095] In this embodiment, an improved adaptive gradient modulation algorithm is used to train the model. This algorithm incorporates the differences in the contribution of different modal data to the prediction results into the model training steps, and combines cosine annealing and regularization. The specific training steps include:

[0096] Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met.

[0097] If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio.

[0098] The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient.

[0099] Based on the modulation coefficients, the updated learning rate, and the updated model gradient, the model parameters corresponding to each modality data are updated, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met.

[0100] In this embodiment, since the parameters of different modes contribute differently to the fault prediction of the distribution ring network cabinet, simply splicing and fusing them will greatly reduce the fault prediction capability. To solve the problem of contribution differences under multi-modal fusion, this invention balances the performance differences of each mode by adjusting the contribution difference ratio, thereby maximizing the optimization capability of the model. Specifically, based on the above three types of frequency domain data, the corresponding ConvLSTM model can obtain the modal data corresponding to different types of frequency domain data. The specific steps are as follows:

[0101] Feature data corresponding to each frequency domain data is obtained from each of the attention-based convolutional long short-term memory networks, and modal weights of each frequency domain data are calculated according to the attention mechanism.

[0102] Based on the feature data and the mode weights, the modal data corresponding to each frequency domain data is obtained, and the modal data includes voltage mode data, current mode data and temperature mode data.

[0103] Specifically, taking the frequency domain data corresponding to the voltage data type as an example, firstly, the feature data extracted by the model is obtained, and then the modal weights of this type of data are calculated according to the attention mechanism. The weights reflect the contribution of the voltage data to the final output result. Here, the modal weights refer to the attention weights in the convolutional long short-term memory network based on the attention mechanism. Since each model is essentially a single-modal model, the attention weights are calculated based on the self-attention mechanism. The specific calculation steps can be referred to the conventional calculation steps of the weights in the attention mechanism, and will not be repeated here.

[0104] After obtaining the feature data and mode weights, the voltage mode data can be obtained based on the following formula:

[0105] In the formula, This represents the voltage mode data obtained in the i-th iteration. This represents the voltage mode weights calculated based on the self-attention mechanism. The weight matrix representing the voltage mode data, b V This represents the bias term for voltage mode data. The voltage feature extraction function θ represents the voltage type in the ConvLSTM model corresponding to the frequency domain data. V The input parameters represent the voltage feature extraction function. Frequency domain data representing voltage type.

[0106] Similarly, current mode data can be expressed by the following formula:

[0107] In the formula, This represents the current mode data obtained in the i-th iteration. This represents the current mode weights calculated based on the self-attention mechanism. The weight matrix representing the current mode data, b C This represents the bias term for current mode data. The current feature extraction function θ represents the current feature extraction function in the ConvLSTM model corresponding to the frequency domain data of the current type. C The input parameters represent the current feature extraction function. Frequency domain data representing current type.

[0108] Temperature modal data can be expressed by the following formula:

[0109] In the formula, This represents the temperature modal data obtained in the i-th iteration. This represents the temperature mode weights calculated based on the self-attention mechanism. The weight matrix representing the temperature modal data, b T This represents the bias term for temperature modal data. The temperature feature extraction function θ in the ConvLSTM model corresponding to the frequency domain data of temperature type T This represents the input parameters of the temperature feature extraction function. Frequency domain data representing temperature type, where t represents the current time step.

[0110] Based on the aforementioned multimodal data, and addressing the differences in optimization across different modalities, this invention monitors the differences in the contribution of each modality to the learning objective and adaptively modulates the gradient of each modality. Performance differences are measured by calculating the contribution difference ratio. Specifically, the modal contribution ratio is first calculated, and then the modulation coefficients of each modal data are calculated based on the modal contribution ratio. The modal contribution ratio is determined by the proportional relationship between each modal data point and the remaining modal data points, as shown in the following formula:

[0111] In the formula, Indicates the modal contribution ratio. Let u represent the modal data obtained in the i-th iteration, u∈{V,C,T}, where V, C, and T represent voltage, current, and temperature, respectively. t This represents a random mini-batch of size m selected at step t, where i represents the iteration number and t represents the number of iteration steps.

[0112] This invention adaptively modulates the gradient by monitoring the aforementioned contribution ratios. For the three modes of voltage, current, and temperature, the modulation coefficients are calculated as follows:

[0113] In the formula, α represents the modulation coefficient of the u-mode data, and α represents the hyperparameter controlling the modulation degree. It should be noted that α here is a hyperparameter based on different modal data. This value can be the same or different for different modal data. The specific value can be flexibly set according to the actual situation.

[0114] To enhance the model's generalization ability, this invention integrates cosine annealing learning rate adjustment and L2 regularization into the existing adaptive gradient modulation algorithm. Cosine annealing gradually reduces the learning rate, ensuring the model converges to a better solution in the later stages of optimization. L2 regularization, on the other hand, limits the infinite increase of model parameters, preventing overfitting. Based on this, the original adaptive learning method is combined with both, resulting in a more stable and robust optimization process.

[0115] The formula for adjusting the learning rate using cosine annealing is as follows:

[0116] Where, η t η is the learning rate for the current step. min It is the minimum value of the learning rate, η. max t is the maximum learning rate, t is the current iteration step, and T is the total number of iterations.

[0117] The standard formula for updating encoder parameters is:

[0118] in, It is an unbiased estimate of the complete gradient, i.e., the model gradient, where η represents the learning rate. B t It is a random mini-batch of size m selected at step t, and At point x, relative to B t The gradient.

[0119] After incorporating the regularization term into the model's gradient, its gradient formula can be expressed as:

[0120] Where λ is the weight of L2 regularization.

[0121] Then, after updating the regular parameters based on the updated learning rate, updated gradient, and modulation coefficients, the parameter update formula is obtained as follows:

[0122] In the formula, This represents the model parameters of the model corresponding to the input data of type u at step t, where u∈{V,C,T}, and V, C, and T represent voltage, current, and temperature, respectively.

[0123] In this embodiment, a contribution difference ratio is designed to balance the performance differences of each mode, addressing the differences in the contribution of parameters from different modes to fault prediction in distribution ring main units. This solves the problem of contribution differences under multi-modal fusion and effectively improves the model's optimization capability. Based on the above structure, the fault detection model provided by this invention also has scalability. For the needs of multiple input modes, as long as the differences in modal contribution are modified accordingly, information from multiple modes can be fused for learning. This allows the model to be widely applied in different distribution ring main unit scenarios, exhibiting high flexibility and scalability.

[0124] This invention utilizes convolutional long short-term memory networks to capture the temporal features of multi-source data, extracts the interrelationships between data points, and integrates information from different data sources into a comprehensive data volume through a multimodal fusion learning mechanism. This provides more comprehensive data support for accurately determining the fault status of distribution ring main units. By combining ConvLSTM networks and multimodal fusion learning, this invention constructs a fault detection model that can deeply mine the features of time-series data and comprehensively integrate multi-source information. This model is not only flexible in application but also has high prediction accuracy, making it highly practical.

[0125] In a preferred embodiment, to enable maintenance personnel to quickly and timely obtain monitoring data from the distribution ring network cabinet, the present invention also provides remote monitoring capabilities. Data is displayed through an intuitive monitoring interface dashboard, including real-time operational data charts, fault logs, and system status indicators. Modern user interface design language and user experience principles are used to ensure ease of use. Data visualization technologies such as charts and heatmaps are employed to make complex data and key performance indicators readily apparent.

[0126] Furthermore, this invention adds an alarm function, assessing the urgency of a fault based on the probability output by the model. For example, when the probability of a fault occurrence exceeds a threshold, an alarm is issued. In addition, to ensure the stable operation of the power system, a preferred embodiment also includes an emergency response mechanism, the specific steps of which include:

[0127] An emergency plan library containing multiple emergency plans is pre-built, and the priority of each emergency plan is set according to the scope of the failure's impact and the urgency of the failure;

[0128] Based on the alarm level indicated by the alarm, the corresponding priority emergency plan is automatically selected and executed from the emergency plan library.

[0129] In this embodiment, multiple emergency response plans and procedures are pre-defined to build an emergency response plan library, enabling the automatic invocation of these plans when a specific fault occurs. These procedures can be prioritized based on the scope and urgency of the fault. These automated measures include disconnecting power, rerouting, or activating backup systems to mitigate the impact of the fault. Then, based on different levels of alarm alerts, the corresponding priority emergency response plan is automatically invoked to ensure the stability of the power system operation.

[0130] In another preferred embodiment, the present invention also employs Bayesian optimization technology to automatically adjust the hyperparameters of the fault detection model. By constructing a probabilistic model of hyperparameters and model performance and using a Gaussian process, not only is effective global optimization of parameters achieved, improving fault detection capabilities, but an adaptive mechanism is also introduced to iteratively update parameters based on performance feedback. This intelligent parameter optimization process significantly reduces the need for manual adjustments. The present invention also combines Bayesian optimization with an automated monitoring and control system to achieve dynamic adjustment of resource allocation. This integrated method can not only intelligently allocate resources based on real-time load conditions, but also optimize the economy of distribution ring network cabinets without affecting service quality, while considering energy consumption and advocating for a more environmentally friendly operating method. Using Bayesian optimization can more accurately estimate the posterior distribution of model parameters, gradually achieving high generalization capability. Bayesian optimization is used to automatically optimize system load management by intelligently allocating tasks among different nodes to optimize load balancing. This technology also has the ability to quickly reallocate resources when facing load fluctuations or deteriorating node performance, ensuring optimal overall system performance while predicting and preventing potential overload problems. By leveraging Bayesian optimization to achieve this advanced automated decision support in system operation and maintenance, the stability and response efficiency of power distribution ring main units are significantly improved. Furthermore, this invention establishes a mechanism for regularly evaluating system performance and security. Through periodic system assessments and upgrades, it ensures that all components remain above industry standards, and system upgrades are performed based on the assessment results, including software updates, hardware replacements, or the addition of new features to adapt to new business requirements.

[0131] This embodiment provides an intelligent fault detection method for distribution ring main units. The invention optimizes the model training process by designing a convolutional long short-term memory network based on multimodal fusion and adaptively adjusting the contribution differences of each mode, thereby improving the model's prediction accuracy. Furthermore, the invention utilizes a constructed remote monitoring platform, enabling maintenance personnel to monitor the key performance indicators of the distribution ring main unit in real time and respond promptly to system-generated early warning and alarm signals, thus improving operational efficiency and system response speed, and further ensuring the safety and stability of the power system operation.

[0132] Please refer to Figure 5. Based on the same inventive concept, the second embodiment of this invention proposes an intelligent fault detection system for a power distribution ring main unit, comprising:

[0133] The determination trigger module 10 is used to acquire the real-time operating data of the power distribution ring network cabinet and determine whether to trigger the fault detection logic based on the fluctuation level of the real-time operating data.

[0134] The fault detection module 20 is used to respond to the triggering of fault detection logic by inputting the real-time running data into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed by a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm.

[0135] The training steps for the fault detection model include:

[0136] Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met.

[0137] If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio.

[0138] The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient.

[0139] The model parameters corresponding to each modality data are updated based on the modulation coefficient, the updated learning rate, and the updated model gradient, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met.

[0140] The alarm notification module 30 is used to generate alarm notifications based on the fault detection results.

[0141] The technical features and effects of the intelligent fault detection system for distribution ring main units proposed in this invention are the same as those of the method proposed in this invention, and will not be repeated here. Each module in the above-mentioned intelligent fault detection system for distribution ring main units can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0142] In summary, the present invention proposes an intelligent fault detection method and system for a distribution ring main unit. The method acquires real-time operating data of the distribution ring main unit and determines whether fault detection logic is triggered based on the fluctuation level of the real-time operating data. In response to triggering the fault detection logic, the real-time operating data is input into a pre-established fault detection model to obtain a fault detection result. The fault detection model is constructed from a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm. An alarm is generated based on the fault detection result. The training steps of the fault detection model include: acquiring historical operating data of the distribution ring main unit and processing the data to obtain frequency domain data; and constructing a fault detection model based on the frequency domain data. A dataset is constructed, and the fault detection model is iteratively trained using the dataset. The process is then repeated to determine if a preset iteration stopping condition has been met. If the iteration stopping condition has not been met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network. The modulation coefficients of each modal data are calculated based on the modal contribution ratio. The learning rate of the model is adjusted using a cosine annealing algorithm, and a regularization term is added to the model gradient to obtain an updated learning rate and an updated model gradient. Based on the modulation coefficients, the updated learning rate, and the updated model gradient, the model parameters corresponding to each modal data are updated, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met. The invention optimizes the model's training process by designing a convolutional long short-term memory network based on multimodal fusion and adaptively adjusting the contribution differences of each modality, effectively improving the model's prediction accuracy. Furthermore, the fault detection model provided by this invention has scalability; for the needs of multiple input modalities, as long as the modal contribution differences are modified accordingly, information from multiple modalities can be fused for learning. This allows the model to be widely applied in different power distribution ring main unit scenarios, exhibiting high flexibility and scalability, further ensuring the safety and stability of power distribution ring main unit operation.

[0143] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0144] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for intelligent fault detection of a power distribution ring main unit, characterized in that, include: Obtain real-time operating data of the power distribution ring network cabinet, and determine whether to trigger fault detection logic based on the fluctuation level of the real-time operating data; In response to triggering the fault detection logic, the real-time running data is input into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed by a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm. Based on the fault detection results, an alarm notification is generated; The training steps of the fault detection model include: Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met. If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio. The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient. Based on the modulation coefficients, the updated learning rate, and the updated model gradient, the model parameters corresponding to each modality data are updated, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met.

2. The method for intelligent detection of faults in a power distribution ring main unit according to claim 1, characterized in that, The fault detection model includes multiple attention-based convolutional long short-term memory networks and a multimodal fusion module, which includes a linear addition layer and an output layer. The number of attention-based convolutional long short-term memory networks is consistent with the number of data types of the frequency domain data. The input data of each attention-based convolutional long short-term memory network is the corresponding type of frequency domain data. The input data of the multimodal fusion module is the output data of each attention-based convolutional long short-term memory network. The output data of the multimodal fusion module is the fault detection result.

3. The method for intelligent detection of faults in a power distribution ring main unit according to claim 1, wherein, The steps of acquiring historical operating data of the power distribution ring network cabinet and processing the data to obtain frequency domain data include: Obtain historical operating data of the power distribution ring main unit, including voltage data, current data, and temperature data; The historical operational data is preprocessed, including missing value imputation, outlier handling, data cleaning, and standardization. The historical running data after data preprocessing is converted into corresponding frequency domain data using a time-frequency conversion algorithm, which includes a fast Fourier transform algorithm based on a mixed basis and a continuous wavelet transform algorithm.

4. The method for intelligent fault detection of power distribution ring main unit according to claim 2, characterized in that, The step of obtaining the modal data corresponding to each frequency domain data based on the output of the convolutional long short-term memory network includes: Feature data corresponding to each frequency domain data is obtained from each of the attention-based convolutional long short-term memory networks, and modal weights of each frequency domain data are calculated according to the attention mechanism. Based on the feature data and the mode weights, the modal data corresponding to each frequency domain data is obtained, and the modal data includes voltage mode data, current mode data and temperature mode data.

5. The method for intelligent detection of faults in a ring main unit according to claim 4, characterized in that, The step of calculating the modulation coefficients of each mode data based on the mode contribution ratio includes: Calculate the modal contribution ratio corresponding to each modal data based on the proportional relationship between each modal data and the other modal data; Based on the modal contribution ratio, the modulation coefficients corresponding to each modal data are calculated.

6. The method for intelligent detection of faults in a ring main unit according to claim 4, characterized in that, The voltage modal data is expressed using the following equation: The current modal data is expressed by the following equation: The temperature modal data is expressed using the following equation: In the formulae, u modal data obtained at the i-th iteration, u e {V, C, T}, V, C and T representing voltage, current and temperature, respectively, u modal weights of the t-th step, a weight matrix representing the u-modal data of the t-th step, b u a bias term representing the u-modal data, a feature extraction function representing u-type data, θ u an input parameter of the feature extraction function representing u-type data, This represents the frequency domain data of type u input during the i-th iteration, where i represents the iteration number and t represents the number of iteration steps.

7. The method for intelligent detection of faults in a ring main unit according to claim 5, characterized in that, The modal contribution ratio is expressed by the following equation: In the formulae, a modality contribution ratio indicative of the u-modal data at the t-th step, u modal data obtained at the i-th iteration, u e {V, C, T}, V, C and T represent voltage, current and temperature, respectively, B t random mini-batch selected at the t-th step, i represents the iteration number, and t represents the iteration step number; The modulation factor is expressed by the following equation: In the formulae, α represents the modulation coefficient of the u-mode data, and α represents the hyperparameter controlling the degree of modulation. The model parameters are expressed by the following equations: In the formulae, denotes the model parameters of the model corresponding to the input data of type u at step t, η t is the updated learning rate of the current step, t is the current iteration step, This represents the updated model gradient.

8. The method for intelligent detection of faults in a ring main unit according to claim 1, characterized in that, The step of determining whether to trigger fault detection logic based on the fluctuation level of the real-time operating data includes: Calculate the standard deviation of the real-time running data and determine whether the standard deviation is greater than a preset threshold. If so, trigger the fault detection logic.

9. The method for intelligent detection of faults in a ring main unit of electrical distribution according to claim 1, characterized in that, After the step of generating an alarm notification based on the fault detection result, the method further includes: An emergency plan library containing multiple emergency plans is pre-built, and the priority of each emergency plan is set according to the scope of the failure's impact and the urgency of the failure; Based on the alarm level indicated by the alarm, the corresponding priority emergency plan is automatically selected and executed from the emergency plan library.

10. A fault intelligent detection system for a power distribution ring main unit, characterized in that, include: The triggering module is used to acquire real-time operating data of the power distribution ring network cabinet and determine whether to trigger the fault detection logic based on the fluctuation level of the real-time operating data. The fault detection module is used to respond to the triggering of fault detection logic by inputting the real-time running data into a pre-established fault detection model to obtain the fault detection result. The fault detection model is constructed by a convolutional long short-term memory network based on multimodal fusion and trained using an improved adaptive gradient modulation algorithm. The training steps for the fault detection model include: Historical operating data of the power distribution ring network cabinet is acquired and processed to obtain frequency domain data. A dataset is constructed based on the frequency domain data. The fault detection model is iteratively trained using the dataset, and it is determined whether the preset iteration stop condition has been met. If the iteration stopping condition is not met, the modal data corresponding to each frequency domain data is obtained based on the output of the convolutional long short-term memory network, and the modulation coefficient of each modal data is calculated based on the modal contribution ratio. The learning rate of the model is adjusted by using the cosine annealing algorithm, and a regularization term is added to the model gradient to obtain the updated learning rate and the updated model gradient. The model parameters corresponding to each modality data are updated based on the modulation coefficient, the updated learning rate, and the updated model gradient, and the fault detection model is trained using the updated model parameters until the iteration stopping condition is met. The alarm notification module is used to generate alarm notifications based on the fault detection results.

Images