A power grid DC transmission equipment fault prediction method based on deep learning

Abstract

The application discloses a kind of power grid DC transmission equipment fault prediction methods based on deep learning, comprising the following steps: collecting and pre-processing the multi-source heterogeneous time series monitoring data of power grid DC transmission equipment;Through embedding projection, generate aligned multi-modal fusion sequence representation;Use multi-scale causal Transform encoder to extract deep feature sequence;Prototype memory bank is constructed and initialized;Through joint optimization instance-prototype and prototype-instance comparison loss, collaborative training feature encoder and prototype memory bank;Based on training model processing historical data generates equipment historical state sequence and constructs state transition probability matrix;Based on real-time data matching and state transition prediction, realize intelligent early warning.The application can automatically learn robust equipment state deep features, effectively find early fault signs and unknown operating modes, significantly improve the accuracy, timeliness and generalization ability of fault warning.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring and fault prediction technology, and in particular to a fault prediction method for DC transmission equipment in power grids based on deep learning. Background Technology

[0002] As a crucial component of modern power grids, the reliable operation of core equipment in high-voltage direct current (HVDC) transmission systems directly impacts the safety and stability of the entire power system. To achieve precise operation and maintenance of critical HVDC transmission equipment such as converter valves and transformers, fault prediction technology has become a key research focus in this field. Currently, data-driven fault prediction methods primarily rely on extracting effective features from equipment operation monitoring data and constructing predictive models.

[0003] Existing technical approaches are mainly divided into two categories: physical mechanism-based modeling and data-driven analysis. Physical mechanism-based methods establish mathematical models of the electrical, thermodynamic, and mechanical behavior of equipment, calculating equipment lifespan loss and failure risk based on theoretical formulas. These methods rely on a deep understanding of the equipment's internal mechanisms; the model-building process is complex and struggles to fully encompass the nonlinear effects of multiple coupled factors in actual operation, limiting their predictive ability for sudden failures and early, subtle signs. Data-driven methods, on the other hand, utilize historical equipment operating data, learning data patterns through machine learning algorithms to achieve prediction. Common approaches include using traditional machine learning models such as support vector machines and random forests to classify or regress manually extracted time-domain and frequency-domain features, and using recurrent neural networks for end-to-end modeling of time-series data. However, these solutions generally suffer from the following shortcomings: First, feature extraction is highly dependent on the prior knowledge and experience of domain experts, and manually designed features often fail to capture complex nonlinear correlations and early degradation signals in high-dimensional, multi-source heterogeneous data; second, traditional machine learning models are weak in handling long sequences and global dependencies, while recurrent neural networks have the problem of difficulty in capturing long-term dependencies, resulting in the model's inaccurate characterization of the long-term evolution trend of equipment status; finally, most solutions have insufficient generalization ability in scenarios where labeled fault samples are scarce, making it difficult to adapt to individual differences and diverse fault modes of different equipment.

[0004] Therefore, how to provide a deep learning-based method for predicting faults in DC power transmission equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based fault prediction method for DC transmission equipment in power grids. Addressing the problems of weak generalization ability and insufficient early fault symptom detection in existing technologies due to reliance on human experience and prior knowledge for data feature extraction, this invention proposes a method that automatically extracts deep contextual features from multi-source heterogeneous time-series data using a multi-scale causal Transformer encoder. It also introduces a learnable dynamic prototype memory and a dual-path contrastive loss function to achieve co-evolution of feature learning and unsupervised state clustering, ultimately providing intelligent early warning based on state transition probabilities. This invention possesses the advantages of automatically learning robust and discriminative equipment state features, effectively discovering unknown operating modes and early fault symptoms, and significantly improving early warning accuracy and model generalization ability.

[0006] A method for predicting faults in DC transmission equipment for power grids based on deep learning, according to an embodiment of the present invention, includes the following steps: S1: Collect multi-source heterogeneous time-series monitoring data streams from DC transmission equipment in the power grid, preprocess them, and obtain regularized multi-source time-series data; S2: Input the regularized multi-source temporal data into the embedding and projection network to generate an aligned multimodal fusion sequence representation; S3: Input the aligned multimodal fused sequence representation into a multi-scale causal Transformer encoder to extract and fuse multi-scale contextual features, and output a deep feature sequence; S4: Initialize a prototype memory consisting of K learnable prototype vectors, where the prototype vectors represent different potential operating states of the device; S5: Based on the deep feature sequence and prototype memory, calculate the instance-prototype contrastive loss and the prototype-instance contrastive loss. Iteratively update the parameters of the multi-scale causal Transformer encoder and the prototype memory by jointly optimizing the contrastive loss until the training converges, and obtain the trained feature encoder and the optimized prototype memory. S6: Use the fully trained feature encoder to process historical monitoring data, obtain historical deep feature sequences, and match them with the optimized prototype memory to generate a device historical state sequence; S7: Based on the historical state sequence of the equipment, count the frequency of state transitions and construct a state transition probability matrix; S8: For real-time monitoring data, extract real-time features using a fully trained feature encoder, and based on the matching results with the optimized prototype memory, perform state evaluation and evolution prediction in conjunction with the state transition probability matrix. When the evaluation or prediction results exceed the preset threshold, output a fault warning signal.

[0007] Optionally, S1 specifically includes: S11: Collect electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data, and environmental monitoring data of DC transmission equipment in the power grid; S12: Construct each type of collected monitoring data into a raw time series with timestamps; S13: Perform outlier detection on each raw time series, identify and mark data points that exceed the preset normal range; S14: For data points marked as outliers and missing data points in the original time series, interpolation is performed using nearby valid data points to fill in the missing data points and generate a continuous time series. S15: Perform standardization transformation on the continuous time series corresponding to each type of monitoring data to eliminate differences in different physical dimensions and magnitudes, and generate standardized time series. S16: Align and arrange the standardized time series of all monitored data types according to a unified timestamp benchmark, and combine them to form aligned multi-source time series data.

[0008] Optionally, S2 specifically includes: S21: Obtain the regularized multi-source time series data from step S1, the data including the standardized time series corresponding to the four types of monitoring data: electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data and environmental monitoring data; S22: Generate a timestamp embedding vector for each data point in the standardized time series; the generation process is: converting the timestamp information corresponding to the data point into a continuous vector representation of fixed dimensions; S23: Construct four independent linear projection layers, corresponding to four types of monitoring data: electrical, mechanical, thermodynamic, and environmental; each linear projection layer receives the standardized time series of the corresponding category as input; S24: Input the standardized time series of the electrical monitoring data into the first linear projection layer, input the standardized time series of the mechanical monitoring data into the second linear projection layer, input the standardized time series of the thermodynamic monitoring data into the third linear projection layer, and input the standardized time series of the environmental monitoring data into the fourth linear projection layer; S25: Each linear projection layer independently performs linear transformation and dimensionality reduction on the input time series data, mapping data with different physical dimensions to a shared latent space with a unified dimension, generating four sets of projection feature sequences. S26: Add the four sets of projection feature sequences to the corresponding timestamp embedding vectors generated in step S22 element by element to achieve the fusion of feature and time information; S27: The four sets of feature sequences after time information fusion are spliced ​​together along the feature dimension to form the aligned multimodal fusion sequence representation.

[0009] Optionally, S3 specifically includes: S31: Based on the aligned multimodal fusion sequence representation, an encoder network is constructed, which includes a first causal attention sublayer, a second causal attention sublayer, and a third causal attention sublayer, which are set up in parallel. S32: Synchronously input the aligned multimodal fusion sequence representation into the first causal attention sublayer, the second causal attention sublayer, and the third causal attention sublayer; S33: In the first causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vectors that precede it and belong to the first neighboring time range, and generate the first scale feature vector; S34: In the second causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the second neighboring time range, and generate a second-scale feature vector, where the second neighboring time range is greater than the first neighboring time range. S35: In the third causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the third neighboring time range, and generate a third-scale feature vector, wherein the third neighboring time range is greater than the second neighboring time range. S36: For each time position in the sequence, the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector corresponding to that position are merged, and the merged result is input into a feedforward network for transformation, outputting the final feature vector of that time position; the final feature vectors of all time positions are arranged in the original time order to form a deep feature sequence.

[0010] Optionally, S4 specifically includes: S41: Based on the total number of prototype vectors in the prototype memory, create a set of learnable vectors with the same dimension as the deep feature vectors output in step S3, and the number of vectors is the same as the total number of prototype vectors. S42: Extract batch deep feature sequences from the training dataset, and aggregate the deep feature vectors at all times in the sequence to form a feature set; S43: Perform cluster analysis on the feature set, divide the feature set into multiple categories, the number of categories is the same as the total number of prototype vectors, and calculate the center vector of each category; S44: Assign the calculated values ​​of the center vectors of each category to the learnable vectors respectively; S45: Organize the learned vectors that have been numerically assigned into an ordered set, which constitutes the initial prototype memory.

[0011] Optionally, S5 specifically includes: S51: Extract a batch of deep feature sequences from the training dataset, which contain deep feature vectors at multiple time points; S52: Calculate the similarity between each deep feature vector in the batch sequence and each prototype vector in the prototype memory, and obtain the similarity matrix; S53: Normalize the similarity values ​​of each row in the similarity matrix and convert the row into a probability distribution that describes the probability that the corresponding deep feature vector belongs to each prototype vector. S54: Calculate the first contrastive loss term based on the probability distribution, which prompts each deep feature vector to move closer to its most likely prototype vector. S55: Calculate the second contrastive loss term based on the probability distribution, which prompts each prototype vector to approach the aggregated representation of the deep feature vectors that may belong to it. S56: Add the first contrast loss term and the second contrast loss term to obtain the joint contrast loss; S57: Using the gradient descent algorithm, based on the gradient calculated by the joint contrastive loss, the internal parameters of the multi-scale causal Transformer encoder and the values ​​of the prototype vectors in the prototype memory are adjusted to obtain the trained feature encoder and the updated prototype memory.

[0012] Optionally, S6 specifically includes: S61: Obtain monitoring data of DC transmission equipment in the power grid within a historical time period; S62: Based on the trained feature encoder, process the historical monitoring data to obtain the historical deep feature sequence; S63: Based on the prototype vectors in the updated prototype memory, for each deep feature vector in the historical deep feature sequence, calculate its matching degree with each prototype vector in the prototype memory. S64: For each deep feature vector, select the prototype vector with the highest matching degree, and mark the preset state category corresponding to the prototype vector as the device state at that moment. S65: Arrange all the device states marked at all historical moments in chronological order to form a sequence of device historical state categories.

[0013] Optionally, S7 specifically includes: S71: Based on the device's historical state category sequence; extract all the state categories that have appeared from the device's historical state category sequence to form a state category set; S72: Traverse the state categories at each pair of adjacent time points in the device's historical state category sequence and record the state transition from the previous state category to the next state category; S73: Count the number of times each state category in the set of state categories is transitioned to each state category in the set of state categories when each state category is used as the starting state; S74: Construct a square matrix whose rows and columns are arranged in the order of state categories in the state category set, and the matrix elements are the number of times the state category is transitioned to each state category in the state category set; S75: Generate the state transition probability matrix based on a square matrix.

[0014] Optionally, S8 specifically includes: S81: Collect monitoring data of DC transmission equipment in the power grid at the current time and in the vicinity of historical time to form a real-time data stream; S82: Use the trained feature encoder to process the real-time data stream to obtain a real-time deep feature sequence; S83: Calculate the matching degree between the latest feature vector in the real-time deep feature sequence and each prototype vector in the updated prototype memory; S84: Based on the matching degree, determine the probability that the feature vector at the latest time step belongs to each prototype vector; S85: Determine whether the sum of probabilities of prototype vectors belonging to a predefined set of unhealthy state categories exceeds a preset first threshold. If it does, trigger the first warning. S86: Based on the equipment's historical state category sequence and state transition probability matrix, starting from the state identified at the current moment, calculate the predicted probability of entering the fault state category within a specified future time period; S87: Determine whether the predicted probability exceeds the preset second threshold. If it does, trigger the second warning. S88: When the first or second warning is triggered, a warning signal is sent to the external system.

[0015] The beneficial effects of this invention are: This invention employs a multi-scale causal Transformer encoder to perform deep feature extraction and fusion of multi-source heterogeneous time-series monitoring data from DC transmission equipment in the power grid. This enables automatic learning of complex nonlinear time-series patterns and long-range dependencies, overcoming the limitations of traditional methods that rely on manual feature engineering and significantly improving the discriminative power of feature representation. By introducing a learnable dynamic prototype memory and combining it with a dual-path contrastive loss function of instance-prototype and prototype-instance, a co-evolutionary mechanism of feature encoding and state clustering is constructed. This allows the model to automatically discover and distinguish different health states and fault modes of equipment under unsupervised conditions, effectively improving the model's generalization ability to unknown operating conditions and novel faults. By constructing the historical state sequence of equipment and calculating the state transition probability matrix in the fault early warning decision-making stage, predictive early warning based on state evolution is achieved. This upgrades early warning from a single-moment judgment to an assessment of the equipment's health degradation trajectory, making the early warning more forward-looking and interpretable. As a result, maintenance instructions can be issued earlier and more accurately in the operation and maintenance of DC transmission equipment in the power grid, effectively improving equipment operational reliability and reducing the risk of unplanned outages. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a schematic diagram of the process of a deep learning-based fault prediction method for DC transmission equipment in a power grid proposed in this invention. Fig. 2 This is a schematic diagram of the co-evolutionary structure of feature encoding and deep comparative clustering proposed in this invention. Detailed Implementation

[0017] Combination Figs. 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Fig. 1 and Fig. 2 The present invention provides a method for predicting faults in DC transmission equipment for power grids based on deep learning, comprising the following steps: S1: Collect multi-source heterogeneous time-series monitoring data streams from DC transmission equipment in the power grid, preprocess them, and obtain regularized multi-source time-series data; S2: Input the regularized multi-source temporal data into the embedding and projection network to generate an aligned multimodal fusion sequence representation; S3: Input the aligned multimodal fused sequence representation into a multi-scale causal Transformer encoder to extract and fuse multi-scale contextual features, and output a deep feature sequence; S4: Initialize a prototype memory consisting of K learnable prototype vectors, where the prototype vectors represent different potential operating states of the device; S5: Based on the deep feature sequence and prototype memory, calculate the instance-prototype contrastive loss and the prototype-instance contrastive loss. Iteratively update the parameters of the multi-scale causal Transformer encoder and the prototype memory by jointly optimizing the contrastive loss until the training converges, and obtain the trained feature encoder and the optimized prototype memory. S6: Use the fully trained feature encoder to process historical monitoring data, obtain historical deep feature sequences, and match them with the optimized prototype memory to generate a device historical state sequence; S7: Based on the historical state sequence of the equipment, count the frequency of state transitions and construct a state transition probability matrix; S8: For real-time monitoring data, extract real-time features using a fully trained feature encoder, and based on the matching results with the optimized prototype memory, perform state evaluation and evolution prediction in conjunction with the state transition probability matrix. When the evaluation or prediction results exceed the preset threshold, output a fault warning signal.

[0018] In this embodiment, S1 specifically includes: S11: Collect electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data, and environmental monitoring data of DC transmission equipment in the power grid; S12: Construct each type of collected monitoring data into a raw time series with timestamps; S13: Perform outlier detection on each raw time series, identify and mark data points that exceed the preset normal range; S14: For data points marked as outliers and missing data points in the original time series, interpolation is performed using nearby valid data points to fill in the missing data points and generate a continuous time series. S15: Perform standardization transformation on the continuous time series corresponding to each type of monitoring data to eliminate differences in different physical dimensions and magnitudes, and generate standardized time series. S16: Align and arrange the standardized time series of all monitored data types according to a unified timestamp benchmark, and combine them to form aligned multi-source time series data.

[0019] In this embodiment, S2 specifically includes: S21: Obtain the regularized multi-source time series data from step S1, the data including the standardized time series corresponding to the four types of monitoring data: electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data and environmental monitoring data; S22: Generate a timestamp embedding vector for each data point in the standardized time series; the generation process is: converting the timestamp information corresponding to the data point into a continuous vector representation of fixed dimensions; S23: Construct four independent linear projection layers, corresponding to four types of monitoring data: electrical, mechanical, thermodynamic, and environmental; each linear projection layer receives the standardized time series of the corresponding category as input; S24: Input the standardized time series of the electrical monitoring data into the first linear projection layer, input the standardized time series of the mechanical monitoring data into the second linear projection layer, input the standardized time series of the thermodynamic monitoring data into the third linear projection layer, and input the standardized time series of the environmental monitoring data into the fourth linear projection layer; S25: Each linear projection layer independently performs linear transformation and dimensionality reduction on the input time series data, mapping data with different physical dimensions to a shared latent space with a unified dimension, generating four sets of projection feature sequences. S26: Add the four sets of projection feature sequences to the corresponding timestamp embedding vectors generated in step S22 element by element to achieve the fusion of feature and time information; S27: The four sets of feature sequences after time information fusion are spliced ​​together along the feature dimension to form the aligned multimodal fusion sequence representation.

[0020] In this embodiment, S3 specifically includes: S31: Based on the aligned multimodal fusion sequence representation, an encoder network is constructed, which includes a first causal attention sublayer, a second causal attention sublayer, and a third causal attention sublayer, which are set up in parallel. S32: Synchronously input the aligned multimodal fusion sequence representation into the first causal attention sublayer, the second causal attention sublayer, and the third causal attention sublayer; S33: In the first causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vectors that precede it and belong to the first neighboring time range, and generate the first scale feature vector; S34: In the second causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the second neighboring time range, and generate a second-scale feature vector, where the second neighboring time range is greater than the first neighboring time range. S35: In the third causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the third neighboring time range, and generate a third-scale feature vector, wherein the third neighboring time range is greater than the second neighboring time range. S36: For each time position in the sequence, the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector corresponding to that position are merged, and the merged result is input into a feedforward network for transformation, outputting the final feature vector of that time position; the final feature vectors of all time positions are arranged in the original time order to form a deep feature sequence.

[0021] In this embodiment, S4 specifically includes: S41: Based on the total number of prototype vectors in the prototype memory, create a set of learnable vectors with the same dimension as the deep feature vectors output in step S3, and the number of vectors is the same as the total number of prototype vectors. S42: Extract batch deep feature sequences from the training dataset, and aggregate the deep feature vectors at all times in the sequence to form a feature set; S43: Perform cluster analysis on the feature set, divide the feature set into multiple categories, the number of categories is the same as the total number of prototype vectors, and calculate the center vector of each category; S44: Assign the calculated values ​​of the center vectors of each category to the learnable vectors respectively; S45: Organize the learned vectors that have been numerically assigned into an ordered set, which constitutes the initial prototype memory.

[0022] In this embodiment, S5 specifically includes: S51: Extract a batch of deep feature sequences from the training dataset, which contain deep feature vectors at multiple time points; S52: Calculate the similarity between each deep feature vector in the batch sequence and each prototype vector in the prototype memory, and obtain the similarity matrix; S53: Normalize the similarity values ​​of each row in the similarity matrix and convert the row into a probability distribution that describes the probability that the corresponding deep feature vector belongs to each prototype vector. S54: Calculate the first contrastive loss term based on the probability distribution, which prompts each deep feature vector to move closer to its most likely prototype vector. S55: Calculate the second contrastive loss term based on the probability distribution, which prompts each prototype vector to approach the aggregated representation of the deep feature vectors that may belong to it. S56: Add the first contrast loss term and the second contrast loss term to obtain the joint contrast loss; S57: Using the gradient descent algorithm, based on the gradient calculated by the joint contrastive loss, the internal parameters of the multi-scale causal Transformer encoder and the values ​​of the prototype vectors in the prototype memory are adjusted to obtain the trained feature encoder and the updated prototype memory.

[0023] In this embodiment, S6 specifically includes: S61: Obtain monitoring data of DC transmission equipment in the power grid within a historical time period; S62: Based on the trained feature encoder, process the historical monitoring data to obtain the historical deep feature sequence; S63: Based on the prototype vectors in the updated prototype memory, for each deep feature vector in the historical deep feature sequence, calculate its matching degree with each prototype vector in the prototype memory. S64: For each deep feature vector, select the prototype vector with the highest matching degree, and mark the preset state category corresponding to the prototype vector as the device state at that moment. S65: Arrange all the device states marked at all historical moments in chronological order to form a sequence of device historical state categories.

[0024] In this embodiment, S7 specifically includes: S71: Based on the device's historical state category sequence; extract all the state categories that have appeared from the device's historical state category sequence to form a state category set; S72: Traverse the state categories at each pair of adjacent time points in the device's historical state category sequence and record the state transition from the previous state category to the next state category; S73: Count the number of times each state category in the set of state categories is transitioned to each state category in the set of state categories when each state category is used as the starting state; S74: Construct a square matrix whose rows and columns are arranged in the order of state categories in the state category set, and the matrix elements are the number of times the state category is transitioned to each state category in the state category set; S75: Generate the state transition probability matrix based on a square matrix.

[0025] In this embodiment, S8 specifically includes: S81: Collect monitoring data of DC transmission equipment in the power grid at the current time and in the vicinity of historical time to form a real-time data stream; S82: Use the trained feature encoder to process the real-time data stream to obtain a real-time deep feature sequence; S83: Calculate the matching degree between the latest feature vector in the real-time deep feature sequence and each prototype vector in the updated prototype memory; S84: Based on the matching degree, determine the probability that the feature vector at the latest time step belongs to each prototype vector; S85: Determine whether the sum of probabilities of prototype vectors belonging to a predefined set of unhealthy state categories exceeds a preset first threshold. If it does, trigger the first warning. S86: Based on the equipment's historical state category sequence and state transition probability matrix, starting from the state identified at the current moment, calculate the predicted probability of entering the fault state category within a specified future time period; S87: Determine whether the predicted probability exceeds the preset second threshold. If it does, trigger the second warning. S88: When the first or second warning is triggered, a warning signal is sent to the external system.

[0026] Example 1: To verify the feasibility and technical effectiveness of this invention in practice, a converter valve—a core component of a converter station in a high-voltage direct current (HVDC) transmission project—is used as the specific application object. As a core power conversion device in a HVDC transmission system, the operating status of the converter valve directly affects the safety and stability of the entire system. Under traditional operation and maintenance models, the assessment of the converter valve's health status mainly relies on periodic preventative tests, offline oil chromatography analysis, and the operator's experience-based threshold judgment of some online monitoring parameters. This model has a significant lag, making it difficult to capture early, weak performance degradation signals before a fault occurs, such as slight aging of individual thyristor modules, a slow decline in cooling system efficiency, or gradual deterioration of electrical connection points. Existing technical solutions attempt to introduce data analysis models based on support vector machines or long short-term memory networks. However, due to the diverse sources of monitoring data for converter valves (including valve hall ambient temperature, cooling water inlet and outlet temperatures, valve tower vibration, surge arrester leakage current, etc.), and the different sampling frequencies and physical meanings of each parameter, severe data heterogeneity is caused. At the same time, the number of fault samples inside the valve body is extremely scarce. This makes traditional machine learning methods heavily reliant on expert knowledge in feature engineering. The models are prone to overfitting to a limited number of fault samples, resulting in poor generalization ability and an inability to effectively identify unknown fault precursor patterns.

[0027] To address the technical challenges in the aforementioned practical scenario, the method of this invention was deployed and applied to the production control platform of the converter station. Multi-source monitoring data related to the converter valves were continuously collected from the station's data acquisition and monitoring control system over a period of three years. This data specifically includes electrical quantity data sampled once per second, such as valve-side AC voltage, DC voltage, and valve arrester leakage current; thermodynamic quantity data sampled once per minute, such as the temperature of key temperature measurement points on the valve tower, the inlet and outlet temperatures of the internal cooling water, and the ambient temperature of the valve hall; and mechanical quantity data sampled once every 10 milliseconds, such as the vibration acceleration of the valve tower foundation. The collected raw data constitutes a multi-source heterogeneous time-series data stream.

[0028] According to the steps described in the claims of this invention, the data is preprocessed. For each type of monitoring data time series, outlier detection methods based on statistical distributions are used to identify and process outliers. For missing data caused by sensor communication interruptions, linear interpolation of adjacent valid data is used to fill in the missing data. After anomaly and missing data processing, each type of data is standardized to convert all parameters to the same numerical order of magnitude. The preprocessed data is aligned according to a unified timestamp to form a regularized multi-source time series data.

[0029] The regularized data is input into a projection network. This network consists of a timestamp embedding layer and four independent linear projection layers. The timestamp embedding layer converts the time information of each data point into a fixed-dimensional vector. The four linear projection layers correspond to four types of data: electrical, thermal, mechanical, and environmental, respectively, mapping sequences of different physical dimensions to a shared 128-dimensional latent space. The projected features are added to the corresponding timestamp embedding vectors, and then the feature dimensions are concatenated to generate an aligned multimodal fusion sequence representation.

[0030] This sequence represents the input to a multi-scale causal Transformer encoder. This encoder comprises three parallel causal self-attention sublayers with attention window lengths of 8, 64, and 512 time steps, respectively, to capture short-term fluctuations, medium-term cycles, and long-term trends. After each sublayer operates independently, the output features are concatenated at each time step, then fused and nonlinearly transformed by a feedforward neural network layer, resulting in a 256-dimensional deep feature sequence.

[0031] During model training, the first two years' data were selected from the three-year dataset as the training set. A memory containing 10 learnable prototype vectors with a dimension of 256 was initialized. Initial centroids were calculated using the deep features of the training set via a clustering algorithm to initialize the prototypes. During training, a batch size of 32 data points was used, with a temperature parameter set to 0.1. The Adam optimizer was employed with an initial learning rate of 0.001. For each batch, the cosine similarity matrix between the deep feature vectors and the prototype vectors was calculated, and the soft-assignment probability was obtained using the softmax function. Based on this probability, both instance-prototype contrastive loss and prototype-instance contrastive loss were calculated with a weight ratio of 1:1. Encoder parameters and prototype vectors were jointly optimized via backpropagation. Training consisted of 300 iterations, and the loss function converged to a stable value after approximately 200 iterations.

[0032] After training, the trained model is used to process historical data from the third year to obtain a deep feature sequence, which is then matched with an optimized prototype memory to generate a daily device status category sequence. The frequency of state transitions in this sequence is statistically analyzed to construct a 10×10 state transition probability matrix. During the real-time warning phase, the system continuously receives the latest monitoring data stream and processes it through the same procedure to obtain real-time deep features. The system performs two parallel judgments: first, it calculates the sum of probabilities that the current feature belongs to two predefined prototype categories, "sub-health" and "warning" (a total of four); if this exceeds a threshold of 0.6, an immediate warning is triggered; second, based on the currently identified state, the system uses the state transition probability matrix to predict the cumulative probability of entering a "warning" or "fault" state within the next 24 hours; if this exceeds a threshold of 0.3, a predictive warning is triggered.

[0033] To quantitatively evaluate the effectiveness of this invention, data from the week preceding five planned maintenance events that occurred and were subsequently confirmed during the third year, as well as 25 weeks of normal operation data, were selected as test samples. The method of this invention was compared with traditional long short-term memory network prediction models and threshold alarm methods based on support vector machines. Evaluation metrics included warning accuracy, false alarm rate, false negative rate, average warning lead time, and the model's recall rate under unknown fault modes. Specific comparative data are shown in the table below: Table 1. Performance comparison of different methods in converter valve fault early warning Statistical analysis of Table 1 shows that, over a 30-week testing period, the early warning accuracy of the method presented in this invention reached 94.7%, significantly higher than the 78.2% of the Long Short-Term Memory (LSTM) network model and the 65.5% of the Support Vector Machine (SVM) threshold alarm method. Regarding the false alarm rate, the method presented in this invention had a false alarm rate of only 3.1%, while the LSTM network model had a false alarm rate of 12.5% ​​and the SVM method had a false alarm rate of 9.4%. Most importantly, for three actual performance degradation events (weeks 1, 2, and 3 before planned maintenance), the method presented a successful early warning with no missed warnings, while the traditional method experienced multiple missed warnings. The average early warning lead time provided by this invention reaches 114 hours, allowing ample time for operational decisions. Furthermore, in the later stages of testing, a minor anomaly caused by a novel environmental factor, which had not appeared in the training set, was successfully identified as an "attention" state by the method presented in this invention, thanks to the strong generalization ability granted by its deep contrastive clustering mechanism, demonstrating its good adaptability to unknown fault modes. In summary, this embodiment, through specific data comparison and analysis, fully verifies the significant advantages of the technical solution of the present invention in improving the accuracy, timeliness, and reliability of fault early warning for DC transmission equipment in power grids.

Claims

1. A method for predicting faults in DC transmission equipment for power grids based on deep learning, characterized in that, Specifically, the following steps are included: S1: Collect multi-source heterogeneous time-series monitoring data streams from DC transmission equipment in the power grid, preprocess them, and obtain regularized multi-source time-series data; S2: Input the regularized multi-source temporal data into the embedding and projection network to generate an aligned multimodal fusion sequence representation; S3: Input the aligned multimodal fused sequence representation into a multi-scale causal Transformer encoder to extract and fuse multi-scale contextual features, and output a deep feature sequence; S4: Initialize a prototype memory consisting of K learnable prototype vectors, where the prototype vectors represent different potential operating states of the device; S5: Based on the deep feature sequence and prototype memory, calculate the instance-prototype contrastive loss and the prototype-instance contrastive loss. Iteratively update the parameters of the multi-scale causal Transformer encoder and the prototype memory by jointly optimizing the contrastive loss until the training converges, and obtain the trained feature encoder and the optimized prototype memory. S6: Use the fully trained feature encoder to process historical monitoring data, obtain historical deep feature sequences, and match them with the optimized prototype memory to generate a device historical state sequence; S7: Based on the historical state sequence of the equipment, count the frequency of state transitions and construct a state transition probability matrix; S8: For real-time monitoring data, extract real-time features using a fully trained feature encoder, and based on the matching results with the optimized prototype memory, perform state evaluation and evolution prediction in conjunction with the state transition probability matrix. When the evaluation or prediction results exceed the preset threshold, output a fault warning signal.

2. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11: Collect electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data, and environmental monitoring data of DC transmission equipment in the power grid; S12: Construct each type of collected monitoring data into a raw time series with timestamps; S13: Perform outlier detection on each raw time series, identify and mark data points that exceed the preset normal range; S14: For data points marked as outliers and missing data points in the original time series, interpolation is performed using nearby valid data points to fill in the missing data points and generate a continuous time series. S15: Perform standardization transformation on the continuous time series corresponding to each type of monitoring data to eliminate differences in different physical dimensions and magnitudes, and generate standardized time series. S16: Align and arrange the standardized time series of all monitored data types according to a unified timestamp benchmark, and combine them to form aligned multi-source time series data.

3. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21: Obtain the regularized multi-source time series data from step S1, the data including the standardized time series corresponding to the four types of monitoring data: electrical monitoring data, mechanical monitoring data, thermodynamic monitoring data and environmental monitoring data; S22: Generate a timestamp embedding vector for each data point in the standardized time series; the generation process is: converting the timestamp information corresponding to the data point into a continuous vector representation of fixed dimensions; S23: Construct four independent linear projection layers, corresponding to four types of monitoring data: electrical, mechanical, thermodynamic, and environmental; each linear projection layer receives the standardized time series of the corresponding category as input; S24: Input the standardized time series of the electrical monitoring data into the first linear projection layer, input the standardized time series of the mechanical monitoring data into the second linear projection layer, input the standardized time series of the thermodynamic monitoring data into the third linear projection layer, and input the standardized time series of the environmental monitoring data into the fourth linear projection layer; S25: Each linear projection layer independently performs linear transformation and dimensionality reduction on the input time series data, mapping data with different physical dimensions to a shared latent space with a unified dimension, generating four sets of projection feature sequences. S26: Add the four sets of projection feature sequences to the corresponding timestamp embedding vectors generated in step S22 element by element to achieve the fusion of feature and time information; S27: The four sets of feature sequences after time information fusion are spliced ​​together along the feature dimension to form the aligned multimodal fusion sequence representation.

4. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31: Based on the aligned multimodal fusion sequence representation, an encoder network is constructed, which includes a first causal attention sublayer, a second causal attention sublayer, and a third causal attention sublayer, which are set up in parallel. S32: Synchronously input the aligned multimodal fusion sequence representation into the first causal attention sublayer, the second causal attention sublayer, and the third causal attention sublayer; S33: In the first causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vectors that precede it and belong to the first neighboring time range, and generate the first scale feature vector; S34: In the second causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the second neighboring time range, and generate a second-scale feature vector, where the second neighboring time range is greater than the first neighboring time range. S35: In the third causal attention sub-layer, calculate the attention weight between each feature vector in the sequence and the historical feature vector located before it that belongs to the third neighboring time range, and generate a third-scale feature vector, wherein the third neighboring time range is greater than the second neighboring time range. S36: For each time position in the sequence, the first-scale feature vector, the second-scale feature vector, and the third-scale feature vector corresponding to that position are merged, and the merged result is input into a feedforward network for transformation, outputting the final feature vector of that time position; the final feature vectors of all time positions are arranged in the original time order to form a deep feature sequence.

5. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41: Based on the total number of prototype vectors in the prototype memory, create a set of learnable vectors with the same dimension as the deep feature vectors output in step S3, and the number of vectors is the same as the total number of prototype vectors. S42: Extract batch deep feature sequences from the training dataset, and aggregate the deep feature vectors at all times in the sequence to form a feature set; S43: Perform cluster analysis on the feature set, divide the feature set into multiple categories, the number of categories is the same as the total number of prototype vectors, and calculate the center vector of each category; S44: Assign the calculated values ​​of the center vectors of each category to the learnable vectors respectively; S45: Organize the learned vectors that have been numerically assigned into an ordered set, which constitutes the initial prototype memory.

6. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51: Extract a batch of deep feature sequences from the training dataset, which contain deep feature vectors at multiple time points; S52: Calculate the similarity between each deep feature vector in the batch sequence and each prototype vector in the prototype memory, and obtain the similarity matrix; S53: Normalize the similarity values ​​of each row in the similarity matrix and convert the row into a probability distribution that describes the probability that the corresponding deep feature vector belongs to each prototype vector. S54: Calculate the first contrastive loss term based on the probability distribution, which prompts each deep feature vector to move closer to its most likely prototype vector. S55: Calculate the second contrastive loss term based on the probability distribution, which prompts each prototype vector to approach the aggregated representation of the deep feature vectors that may belong to it. S56: Add the first contrast loss term and the second contrast loss term to obtain the joint contrast loss; S57: Using the gradient descent algorithm, based on the gradient calculated by the joint contrastive loss, the internal parameters of the multi-scale causal Transformer encoder and the values ​​of the prototype vectors in the prototype memory are adjusted to obtain the trained feature encoder and the updated prototype memory.

7. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61: Obtain monitoring data of DC transmission equipment in the power grid within a historical time period; S62: Based on the trained feature encoder, process the historical monitoring data to obtain the historical deep feature sequence; S63: Based on the prototype vectors in the updated prototype memory, for each deep feature vector in the historical deep feature sequence, calculate its matching degree with each prototype vector in the prototype memory. S64: For each deep feature vector, select the prototype vector with the highest matching degree, and mark the preset state category corresponding to the prototype vector as the device state at that moment. S65: Arrange all the device states marked at all historical moments in chronological order to form a sequence of device historical state categories.

8. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71: Based on the device's historical state category sequence; extract all the state categories that have appeared from the device's historical state category sequence to form a state category set; S72: Traverse the state categories at each pair of adjacent time points in the device's historical state category sequence and record the state transition from the previous state category to the next state category; S73: Count the number of times each state category in the set of state categories is transitioned to each state category in the set of state categories when each state category is used as the starting state; S74: Construct a square matrix whose rows and columns are arranged in the order of state categories in the state category set, and the matrix elements are the number of times the state category is transitioned to each state category in the state category set; S75: Generate the state transition probability matrix based on a square matrix.

9. The method for predicting faults in DC transmission equipment for power grids based on deep learning according to claim 1, characterized in that, S8 specifically includes: S81: Collect monitoring data of DC transmission equipment in the power grid at the current time and in the vicinity of historical time to form a real-time data stream; S82: Use the trained feature encoder to process the real-time data stream to obtain a real-time deep feature sequence; S83: Calculate the matching degree between the latest feature vector in the real-time deep feature sequence and each prototype vector in the updated prototype memory; S84: Based on the matching degree, determine the probability that the feature vector at the latest time step belongs to each prototype vector; S85: Determine whether the sum of probabilities of prototype vectors belonging to a predefined set of unhealthy state categories exceeds a preset first threshold. If it does, trigger the first warning. S86: Based on the equipment's historical state category sequence and state transition probability matrix, starting from the state identified at the current moment, calculate the predicted probability of entering the fault state category within a specified future time period; S87: Determine whether the predicted probability exceeds the preset second threshold. If it does, trigger the second warning. S88: When the first or second warning is triggered, a warning signal is sent to the external system.

Images