A remote intelligent operation monitoring method and system for a smart substation

Abstract

The application provides a remote intelligent operation monitoring method and system of a smart substation, relates to the technical field of intelligent operation and maintenance of a substation, and relates to multi-source heterogeneous sensing, deep feature modeling, fault prediction evaluation and model self-optimization.The application collects electrical, environmental and meteorological data through heterogeneous sensors, constructs a structured original data set, carries out time series prediction in combination with a convolution-LSTM model, and outputs a fault probability and a confidence interval by using a Transformer fusion network, so that online early warning and response control are realized, and the application has the self-adaptive updating capacity driven by federated learning.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for substations, and in particular to a remote intelligent operation monitoring method and system for intelligent substations. Background Technology

[0002] Smart substations have widely adopted digital primary equipment, fiber optic Ethernet communication, and standardized protocols to achieve online acquisition and centralized management of electrical and environmental quantities. However, with the continuous expansion of the power grid scale and topology complexity, the operating status exhibits strong dynamics, nonlinearity, and uncertainty. Traditional monitoring methods that rely on experience rules are difficult to meet the real-time and accurate safety operation and maintenance needs.

[0003] The industry is accelerating its transformation towards data-driven and intelligent decision-making. Technologies such as deep fusion of multi-source heterogeneous information, artificial intelligence diagnosis, and edge-cloud collaborative self-learning are being introduced into the fault monitoring system, with the goal of upgrading from passive alarms to proactive prediction and adaptive optimization.

[0004] Existing solutions still suffer from problems such as information silos, low accuracy in identifying complex faults, and the inability to dynamically adjust operation and maintenance strategies based on diagnostic results, resulting in delayed early warnings and high false alarm rates, which restrict the safe and economical operation of smart substations. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a remote intelligent operation monitoring method and system for smart substations. This method not only achieves high-dimensional collaborative processing of multi-source data and high-precision prediction of critical states, but also constructs a fully closed-loop early warning and linkage system with self-learning capabilities, significantly improving the real-time perception, risk prediction, and response decision-making capabilities of smart substations under complex operating conditions.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A remote intelligent operation monitoring method for a smart substation includes:

[0008] Heterogeneous sensors are deployed on primary equipment, auxiliary equipment, and the substation environment to collect electrical parameters, environmental parameters, and external meteorological information in real time. The electrical parameters, environmental parameters, and external meteorological information of different protocols are timestamped to form a structured raw dataset.

[0009] Normalization is performed on the numerical data in the structured original dataset to obtain a first processing result, and word embedding encoding is performed on the text data to obtain a second processing result. The first processing result and the second processing result are then merged into a unified feature vector sequence.

[0010] The unified feature vector sequence is encapsulated and encrypted and sent to the remote monitoring platform via a 5G mobile communication link;

[0011] Based on the time-series prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, multi-layer modeling is performed on the unified feature vector sequence to output the prediction sequence of key operating parameters within a future preset time window.

[0012] The structured original dataset and the predicted sequence of key operating parameters are input into the Transformer state fusion network to generate a joint state vector and output the probability of occurrence of each fault type and the corresponding confidence interval.

[0013] When the probability of any of the aforementioned fault types exceeds a set threshold, an early warning message is automatically generated and control, maintenance, and protection measures are triggered simultaneously.

[0014] The parameters of the time-series prediction model and the Transformer state fusion network are updated using historical monitoring data at preset intervals based on a federated learning mechanism, and the heterogeneous sensor nodes are dynamically managed.

[0015] Preferably, heterogeneous sensors are deployed on the primary equipment, auxiliary equipment, and substation environment of the substation to collect electrical parameters, environmental parameters, and external meteorological information in real time. The electrical parameters, environmental parameters, and external meteorological information from different protocols are then timestamped to form a structured raw dataset, including:

[0016] Establish a sensor deployment mapping table, plan the deployment location, measurement type and unique number of the heterogeneous sensors for primary equipment, auxiliary equipment and station environment of substation respectively, and record the physical coordinates of each heterogeneous sensor;

[0017] The sampling tasks of all heterogeneous sensors are started synchronously through a unified broadcast trigger signal, and a local timestamp is generated for each frame of sampled data to form a synchronous sampling data frame.

[0018] A time calibration mechanism based on a reference clock is adopted. For each heterogeneous sensor numbered k, within a preset calibration period, the time offset between the local timestamp sequence and the main reference clock sequence is calculated, and the original timestamp is corrected to an aligned timestamp. The formula for calculating the aligned timestamp is as follows: Among them, t k (n) represents the local original timestamp of sensor number k in the nth sampling; δ k This represents the time offset between the sensor numbered k and the reference clock within the current calibration cycle, calculated as the median of multiple time differences. c(i) represents the timestamp recorded by the reference clock in the i-th calibration broadcast; The timestamp after alignment is given for sensor k after the nth sampling; N represents the number of times the reference clock and sensor k successfully establish communication and record the timestamp difference within a complete calibration cycle;

[0019] Based on a unified data encapsulation structure, the number, aligned timestamp, and corresponding measurement data triplet are combined. The original structured dataset is organized into a standardized format; where v k (n) represents the measurement data, indicating the physical quantity measured by the sensor numbered k during the nth sampling; this measurement value is an electrical parameter, environmental parameter, or external meteorological information.

[0020] Preferably, normalization processing is performed on the numerical data in the structured original dataset to obtain a first processing result, and word embedding encoding is performed on the text data to obtain a second processing result. The first processing result and the second processing result are then merged into a unified feature vector sequence, including:

[0021] For each numerical feature numbered m in the structured original dataset, calculate the 10th percentile P of the numerical feature within a sliding window. 10,m 50th percentile P 50,m 90th percentile P 90,m , used for dynamic normalization processing;

[0022] Using the normalization formula For each record in the structured dataset, the numerical feature value x m (r) is normalized to obtain the normalized result.

[0023] For each text-class feature label q(r) in the record, a pre-trained word embedding matrix W is used. e The one-hot encoded vector o q(r) The projection is the embedding vector e q(r) =W e o q(r) ;

[0024] Form a column vector x from all the normalized numerical features of the record with index r. norm (r), all text embedding vectors are combined to form a column vector e(r), and a unified feature vector s(r) is generated by vector-level concatenation; the formula for calculating the unified feature vector s(r) is: s(r) = x norm (r)‖e(r);

[0025] Where, x m(r) represents the original value of the numerical feature numbered m in the r-th global record of the structured original dataset; ∈ is a positive integer to avoid division by zero; This represents the normalization result of feature number m at record r; o q(r) e is the one-hot encoded vector that records the q-th type of text features in r; q(r) The embedding vector is used to record the corresponding text features in r.

[0026] Preferably, the unified feature vector sequence is encapsulated and encrypted and sent to the remote monitoring platform via a 5G mobile communication link, including:

[0027] Each feature vector in the unified feature vector sequence is segmented into basic data blocks according to a fixed byte length, and each data block is marked with an index number and a timestamp.

[0028] Each data block is encrypted using the national standard SM4 block cipher algorithm, with a unified key K. s A ciphertext block sequence is generated using a fixed initialization vector IV0; the formula for calculating the ciphertext block sequence is as follows: Wherein, s(r j ) represents the sequence of uniform feature vectors with index r. j eigenvectors; Indicates the use of key K s and the initialization vector IV0 to vector s(r) j Perform SM4 symmetric block encryption; C j For the corresponding ciphertext block sequence;

[0029] The encrypted ciphertext block sequence is encapsulated into a data packet frame structure; the data packet frame structure includes a frame header, a frame body, and a frame trailer; the frame header includes the frame type, the number of data blocks, and a synchronization flag; the frame body is the ciphertext block content; the frame trailer is a CRC checksum; the CRC checksum is used for message integrity verification.

[0030] By utilizing a 5G IoT module to access the public network, a remote monitoring platform address and port based on a multi-access edge computing mode are configured. The data packet frame structure is transmitted in an orderly manner through the TCP / IP protocol, and the transmission confirmation response is recorded and resuming interrupted transmission is supported.

[0031] Preferably, based on a time-series prediction model composed of a cascaded one-dimensional convolutional layer and a two-layer long short-term memory neural network, multi-layer modeling is performed on the unified feature vector sequence to output a predicted sequence of key operating parameters within a future preset time window, including:

[0032] A fixed-length sliding window is used to slice the uniform feature vector sequence to obtain the input tensor v(t), and each time step contains all feature dimensions;

[0033] Using a sequence of one-dimensional convolution kernels of length F Perform convolution operations and obtain the convolution feature sequence through a linear rectified activation function;

[0034] The convolutional feature sequences are sequentially input into the first layer of the Long Short-Term Memory (LSTM1) network to extract short-term and long-term temporal dependencies.

[0035] The hidden state sequence of LSTM1 is input into the second layer of the Long Short-Term Memory network LSTM2 to generate a deep temporal representation;

[0036] A linear transformation is performed on the current hidden state of the LSTM2, and the prediction vector is output through the Sigmoid activation function, as shown in the following formula:

[0037]

[0038] Where v(t) is the input tensor of the unified feature vector sequence at time t; Φ p Γ is the p-th convolution weight matrix in the one-dimensional convolution kernel sequence; F is the convolution kernel length; ReLU is the linear rectified activation function; LSTM1 and LSTM2 represent the first and second layers of the long short-term memory network operations, respectively; Γ is the output mapping matrix; β is the output bias vector; σ is the sigmoid activation function. This is the prediction vector for the key operating parameter τ relative to the current time.

[0039] Preferably, the structured original dataset and the predicted sequence of key operating parameters are input into the Transformer state fusion network to generate a joint state vector, and the probability of occurrence of each fault type and the corresponding confidence interval are output, including:

[0040] Under the same time index r, the normalized numerical-text feature vector x(r) and the corresponding key running parameter prediction vector in the original structured dataset are compared. Perform vector-level concatenation to obtain the concatenated input vector.

[0041] To concatenate the input vector sequence Superimposed learnable position embedding matrix P enc This forms the encoding matrix X = U + P enc ;

[0042] The encoding matrix X is sequentially fed into an L-layer multi-head self-attention Transformer encoder, and the self-attention output within each layer is calculated using the following formula:

[0043] Q = XW Q K = XWK V = XW V

[0044] Take the first row vector h of the output matrix of the last encoder layer cls As a joint state vector, the joint state vector is combined through a linear mapping-Softmax to obtain the fault probability vector p = softmax(W). o h cls +b o );

[0045] During the inference phase, M random deactivated forward propagations are performed on the same input to obtain a probability sample set. The confidence interval for each fault type f is estimated using the mean-variance method, and the formula is as follows:

[0046]

[0047] CI f =[μ f -Z 0.975 σ f ,μ f +Z 0.975 σ f ];

[0048] Where U is a matrix composed of all u(r) in chronological order; P enc W is the position embedding matrix; Q W K W V d is the query, key, and value mapping matrix; d is the model hidden dimension; W o ,b o The classification mapping weights and biases are defined; p is the fault probability vector obtained from a single inference step. μ represents the probability of fault type f in the m-th random inactivation inference. f and Z represents the mean and variance of the probability of fault type f, respectively; 0.975 =1.96 is the 95% confidence coefficient; CI f is the 95% confidence interval for fault type f; T is the length of the joint input vector sequence; M represents the number of samples for forward propagation performed using the random deactivation mechanism during the inference phase.

[0049] Preferably, when the probability of any of the aforementioned fault types exceeds a set threshold, an early warning message is automatically generated and control, maintenance, and protection measures are triggered simultaneously, including:

[0050] Establish a mapping table between fault types and response levels, set response thresholds for various fault types, and classify them into three levels according to severity: general alarms, important alarms, and emergency alarms.

[0051] The probability of each type of fault output by the state fusion network is compared item by item. If any fault probability is higher than its corresponding alarm threshold, it is determined to be a fault warning trigger event.

[0052] Based on the matched alarm level, structured early warning information is generated in the monitoring platform; the early warning information includes the fault type, probability of occurrence, estimated time, scope of impact, and response suggestions.

[0053] The warning information is synchronously sent to the local substation control device through the communication gateway, and the corresponding response strategy is executed according to the alarm level.

[0054] Preferably, the response strategy includes:

[0055] General alarm trigger status records and manual inspection prompts;

[0056] Critical alarms trigger partition load adjustments and planned maintenance requests;

[0057] Emergency alarms directly trigger fault isolation, backup power switching, and protection actions.

[0058] Preferably, the parameters of the time-series prediction model and the Transformer state fusion network are updated using historical monitoring data at preset intervals based on a federated learning mechanism, and the heterogeneous sensor nodes are dynamically managed, including:

[0059] A fixed federated training cycle is set in the remote monitoring platform. At the beginning of each cycle, the initial parameters of the model are sent from the central coordinator to each participating site.

[0060] Historical monitoring data stored in each substation's local node is retrieved and used to train the time-series prediction model and the local copy of the Transformer state fusion network, respectively, and the local model update results are recorded.

[0061] The parameter updates of the local model are sent back to the remote monitoring platform, where the central coordinating unit performs parameter aggregation without aggregating the original data, thus completing the global model update.

[0062] The updated model is replayed for verification and performance evaluation. If the evaluation results meet the target conditions of improved accuracy and reduced false alarm rate, the new model is activated for subsequent prediction and fusion tasks.

[0063] After the model update is completed, the sampling frequency, reporting period and power consumption configuration of the heterogeneous sensor are dynamically adjusted based on the feature redundancy and sampling contribution feedback from the model, so as to achieve adaptive optimization management of the nodes of the heterogeneous sensor.

[0064] A remote intelligent operation monitoring system for a smart substation includes:

[0065] The heterogeneous sensing and acquisition unit is used to deploy heterogeneous sensors on the primary equipment, auxiliary equipment and substation environment of the substation, so as to collect electrical parameters, environmental parameters and external meteorological information in real time through the heterogeneous sensors, and to align the electrical parameters, environmental parameters and external meteorological information of different protocols with timestamps to form a structured raw dataset.

[0066] The feature encoding processing unit is used to perform normalization processing on the numerical data in the structured original dataset to obtain a first processing result, and to perform word embedding encoding on the text data to obtain a second processing result, and to merge the first processing result and the second processing result into a unified feature vector sequence;

[0067] An encrypted transmission unit is used to encapsulate and encrypt the unified feature vector sequence and send it to a remote monitoring platform via a 5G mobile communication link.

[0068] The temporal modeling and prediction unit is used to perform multi-layer modeling on the unified feature vector sequence based on a temporal prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, and output the prediction sequence of key operating parameters within a future preset time window.

[0069] The state fusion evaluation unit is used to input the structured original dataset and the predicted sequence of key operating parameters into the Transformer state fusion network, generate a joint state vector, and output the probability of occurrence of each fault type and the corresponding confidence interval.

[0070] The early warning linkage control unit is used to automatically generate early warning information and simultaneously trigger control, maintenance and protection measures when the probability of any of the aforementioned fault types exceeds a set threshold.

[0071] The model adaptive optimization unit is used to update the parameters of the time-series prediction model and the Transformer state fusion network based on a federated learning mechanism using historical monitoring data at a preset period, and to dynamically manage the heterogeneous sensor nodes.

[0072] The present invention discloses the following technical effects:

[0073] (1) By deploying heterogeneous sensors in primary equipment, auxiliary equipment and environment of substations and introducing a unified timestamp alignment mechanism, this invention realizes the fusion collection and structured coding of multiple types of operational information such as electrical, environmental and meteorological information, breaks through the bottleneck of isolated perception and inconsistent data format in traditional monitoring systems, and significantly improves data integrity and timeliness.

[0074] (2) The present invention adopts a time-series prediction structure jointly modeled by convolution and two-layer long short-term memory network, which fully extracts local patterns and global dependencies in historical operation features, making the prediction of key operation parameters more accurate and the trend judgment more sensitive, and enhancing the ability to predict sudden abnormal states in advance.

[0075] (3) By introducing the Transformer state fusion network, the present invention jointly models real-time sensing data and prediction results, and outputs multiple types of fault probabilities with confidence intervals, thereby achieving accurate identification of complex abnormal states and quantitative expression of uncertainty, effectively reducing the risk of false alarms and missed alarms.

[0076] (4) This invention uses a federated learning mechanism to periodically iterate and optimize the core model, and combines the contribution evaluation of sensor nodes to achieve adaptive management, thus constructing a closed-loop intelligent monitoring system with the ability to "collect data, learn, and optimize simultaneously", which significantly improves the generalization ability and stability in multiple scenarios and working conditions. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0079] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0080] 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, and 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.

[0081] The purpose of this invention is to provide a remote intelligent operation monitoring method and system for smart substations. It not only realizes high-dimensional collaborative processing of multi-source data and high-precision prediction of critical states, but also constructs a fully closed-loop early warning and linkage system with self-learning capabilities, which significantly improves the real-time perception, risk prediction and response decision-making capabilities of smart substations under complex operating conditions.

[0082] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, the present invention provides a remote intelligent operation monitoring method for intelligent substations, comprising:

[0084] Step 100: Deploy heterogeneous sensors on primary equipment, auxiliary equipment and substation environment of substation to collect electrical parameters, environmental parameters and external meteorological information in real time through heterogeneous sensors, and align the timestamps of electrical parameters, environmental parameters and external meteorological information of different protocols to form a structured raw dataset.

[0085] Step 200: Perform normalization on the numerical data in the structured original dataset to obtain the first processing result, and perform word embedding encoding on the text data to obtain the second processing result. Then, merge the first processing result and the second processing result into a unified feature vector sequence.

[0086] Step 300: Encapsulate and encrypt the unified feature vector sequence and send it to the remote monitoring platform via the 5G mobile communication link;

[0087] Step 400: Based on the time series prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, multi-layer modeling is performed on the unified feature vector sequence to output the prediction sequence of key operating parameters within the future preset time window.

[0088] Step 500: Input the structured original dataset and the predicted sequence of key operating parameters into the Transformer state fusion network to generate a joint state vector and output the probability of occurrence of each fault type and the corresponding confidence interval.

[0089] Step 600: When the probability of any fault type occurring exceeds the set threshold, an early warning message is automatically generated and control, maintenance, and protection measures are triggered simultaneously.

[0090] Step 700: Using historical monitoring data at preset intervals, update the parameters of the time series prediction model and the Transformer state fusion network based on the federated learning mechanism, and dynamically manage heterogeneous sensor nodes.

[0091] In step 100 of this embodiment, various types of heterogeneous sensors are deployed for the primary equipment (such as circuit breakers, instrument transformers, and busbars), auxiliary equipment (such as batteries, air conditioners, and control cabinets), and the substation environment (such as outdoor switch areas, main control rooms, and ventilation areas). These sensors include, but are not limited to, current sensors, voltage sensors, temperature and humidity sensors, gas detectors, vibration sensors, and meteorological monitoring instruments. This embodiment establishes a sensor deployment mapping table, configuring each sensor with its measurement object, sensing type, unique identifier, and physical installation location coordinates. This mapping table is used for accurate identification and location mapping of sensors during subsequent data processing.

[0092] Specifically, in the data acquisition phase of step 100, this embodiment broadcasts a unified trigger signal to instruct all heterogeneous sensors to simultaneously start sampling tasks and generate a timestamp corresponding to the current sampling frame locally. To achieve data alignment between sensors with different communication protocols, this embodiment constructs a time calibration mechanism based on a reference clock. Within a preset calibration period, it records the local timestamp sequence reported by each numbered sensor and compares it with the broadcast time of the reference clock, calculating the median time difference as the time offset value of the sensor, thereby converting the original timestamp into a globally aligned timestamp. Based on this, this embodiment organizes each sampling record into a triplet of "sensor number - aligned timestamp - measurement data" and uniformly encapsulates it into a structured raw dataset. The measurement data may include electrical parameters (such as phase current, bus voltage, power factor), environmental parameters (such as temperature, humidity, smoke concentration), or meteorological information (such as wind speed, air pressure, irradiance), etc., to support subsequent feature processing and state modeling.

[0093] In this embodiment, for the numerical data in the structured original dataset described in step 200, quantile normalization is used for standardization to improve the model's robustness to outliers and distribution changes. Specifically, for each numerical feature numbered m, the 10th percentile, 50th percentile (median), and 90th percentile are calculated within a sliding window and used as the low-order, median, and high-order values ​​of the feature in the current time period, respectively. Subsequently, the original feature value numbered m in the r-th record of the global record is normalized according to the method of "current value minus median, then divided by the difference between the high-order and low-order values" to obtain the first standardized processing result. A slightly positive integer is introduced during normalization to prevent the risk of division by zero. The normalized multidimensional numerical features are uniformly organized into a column vector at record r, and the corresponding record index is retained.

[0094] Meanwhile, in step 200 of this embodiment, for textual features (such as equipment status descriptions, fault type labels, etc.) present in the structured original dataset, each text value is first converted into a one-hot encoded vector, and then the one-hot vector is embedded and mapped using the loaded pre-trained word embedding matrix to obtain the corresponding low-dimensional continuous embedding vector, which serves as the second processing result. The semantic vectors of each dimension after text embedding are also organized into column vectors. Finally, in this embodiment, the normalized numerical feature column vector corresponding to each record is concatenated with the embedded text feature column vector according to dimension to form a unified feature vector, and a unified feature vector sequence is formed as the input for downstream modeling. This preprocessing method not only preserves the relative fluctuation characteristics of numerical information but also integrates the semantic relationships of textual information, significantly improving the ability of subsequent models to express temporal patterns and multimodal features.

[0095] In this embodiment, for the unified feature vector sequence described in step 300, to ensure the stability and data security of transmission over the wireless link, the feature vector is first sliced ​​according to a fixed byte length, dividing each complete feature vector into multiple basic data blocks, and assigning a unique index number and generating a corresponding sampling timestamp for each data block. This embodiment preferably uses a data length alignment method for segmentation to ensure that each data block facilitates subsequent encryption processing and data encapsulation.

[0096] In step 300, this embodiment performs group encryption on each of the above data blocks based on the SM4 symmetric encryption algorithm, a national commercial cryptographic algorithm standard. A uniformly set key and initialization vector are used to ensure the confidentiality and tamper resistance of the data during public network transmission. Each encrypted data block constitutes a ciphertext block, and the sequence of all ciphertext blocks forms the ciphertext set to be transmitted. To ensure the integrity of the transmission structure and protocol compatibility, this embodiment encapsulates the ciphertext block sequence into a data packet frame structure. The frame structure includes three parts: a frame header, a frame body, and a frame trailer. The frame header contains metadata such as frame type identifier, number of data blocks, and synchronization flags. The frame body stores the ciphertext data content, and the frame trailer appends a checksum generated based on the Cyclic Redundancy Check (CRC) algorithm for the receiving end to verify message integrity.

[0097] In the transmission phase of step 300, this embodiment achieves public network access through a 5G IoT communication module and configures the receiving address and target port of the remote monitoring platform based on Multi-Access Edge Computing (MEC) mode. The encapsulated message frames are transmitted in an ordered, fragmented manner using the TCP / IP protocol, and an acknowledgment response is received after each data segment is sent to ensure the stability of the transmission link and the success rate of data delivery. Simultaneously, this embodiment supports a breakpoint resumption mechanism, which can automatically identify the interruption location and resume transmission from the interruption point after communication interruption, enhancing the robustness and availability of the system in unstable network environments.

[0098] In step 400 of this embodiment, to perform deep modeling of the unified feature vector sequence to achieve forward prediction of key operating parameters, a time-series prediction model composed of a cascaded one-dimensional convolutional layer and a two-layer long short-term memory neural network (LSTM) is constructed. First, a fixed-length sliding window is used to segment the unified feature vector sequence. Each consecutive time step within the sliding window serves as an input sample, forming a three-dimensional input tensor, where each time step contains the complete feature dimension. This window mechanism is used to capture the dynamic evolution features within a local time period, while simultaneously increasing the number of training samples and the model's generalization ability.

[0099] In step 400 of this embodiment, a one-dimensional convolution operation is applied to the input tensor. A sequence of one-dimensional convolution kernels of length F is used for time-step feature extraction, and the convolutional feature sequence is obtained after activation by a Rectified Linear Unit (ReLU) function. This convolution process is used to extract local change trends and pattern edge information, reducing feature redundancy. The resulting convolutional feature sequence is sequentially input into the first-layer LSTM network to model short-term and medium-term dependencies. The hidden state of the first-layer LSTM is then passed to the second-layer LSTM network for further processing to uncover deeper long-term correlations. Finally, this embodiment performs a linear transformation on the hidden state of the second-layer LSTM at the current time step and connects it to a Sigmoid activation function, outputting the predicted value of the future key operating parameters corresponding to the current time point. The prediction results can cover multiple target variables, such as load current, voltage fluctuations, and ambient temperature trends, to drive subsequent state fusion and fault prediction processing. This modeling structure fully combines the advantages of local feature extraction from convolution with the temporal memory capability of LSTM, achieving accurate modeling and forward-looking judgment of nonlinear, long-dependency state evolution processes without relying on manual feature engineering.

[0100] In step 500 of this embodiment, a Transformer state fusion network is constructed to integrate the multi-source observation state and prediction sequence at the current moment to identify and assess the probability of various operational faults. Specifically, under the premise of consistent global record index, this embodiment merges the normalized numerical features and word-embedded text features in the structured original dataset into a multi-dimensional vector, and concatenates it with the prediction vector of key operational parameters output by the time-series prediction model at the same moment to form a joint input vector. In this way, the state combination of "current observation value + future prediction value" is integrated, improving the timeliness and foresight of abnormal trend identification.

[0101] Subsequently, in step 500 of this embodiment, the aforementioned joint input vector sequence is assembled into an input matrix in chronological order, and a learnable position embedding matrix is ​​superimposed to form a time-aware encoding matrix. This encoding matrix is ​​then input into a multi-layer Transformer encoder, each layer consisting of a multi-head self-attention mechanism and a feedforward network. In the attention mechanism, query (Q), key (K), and value (V) vectors are calculated respectively, and a weighted aggregation operation is performed to extract the correlation influence of key positions in the sequence on the current state. This structure possesses good modeling capabilities for long-range dependencies and is suitable for causal correlation analysis between long-term features in substation operation.

[0102] After completing the L-layer Transformer encoding, this embodiment extracts the first row vector from the output encoding matrix of the last layer as the joint state vector, and obtains the predicted probability distribution of various operational faults through a combination of linear mapping and the Softmax function. This prediction result not only covers common fault types (such as abnormal bus temperature rise, load imbalance, and switch tripping), but also exhibits good multi-class robustness. To quantify the uncertainty of the model output, this embodiment employs a dropout strategy during the inference phase to perform multiple forward propagations on the same input, recording the probability output of each fault type.

[0103] In the final step of step 500, this embodiment statistically analyzes the probability sample set for each fault type, calculates its mean and variance, and estimates the confidence interval range based on the 95% confidence coefficient of the standard normal distribution. This is used to assist maintenance personnel in judging the credibility of their decisions. The mechanism for generating this confidence interval enhances the interpretability of the model, reduces the risk of misjudgment due to data disturbances or feature anomalies, and improves the practical reliability of the entire monitoring system. The above fusion strategy not only improves the accuracy of fault identification but also achieves quantifiable uncertainty assessment, forming a composite decision-making capability of "current state + trend prediction + probability confidence."

[0104] In step 600 of this embodiment, to ensure that corresponding countermeasures can be taken promptly according to the severity level after a fault risk is detected, a mapping table between fault types and response levels is first established. This mapping table presets probability thresholds corresponding to various fault types (such as main bus overload, switch failure, three-phase imbalance, etc.) and classifies response levels into three categories—general alarms, important alarms, and emergency alarms—based on the potential equipment damage, system downtime risk, and safety hazards caused by the fault. Each response level is bound to a specific processing strategy and warning priority for subsequent control decisions.

[0105] During operation, this embodiment monitors and compares the occurrence probability of each fault type output by the Transformer state fusion network in real time. If the predicted probability of a certain fault type exceeds the response threshold set in the mapping table, a fault warning event is immediately triggered. After the system identifies the fault type and its corresponding alarm level, it enters the automatic response phase to ensure that fault handling is timely and targeted, avoiding safety hazards caused by delays in manual judgment.

[0106] To ensure that maintenance personnel can promptly grasp fault details and handling suggestions, this embodiment automatically generates structured early warning information in the remote monitoring platform when an early warning is triggered. This early warning information includes: fault type name, predicted probability value, expected occurrence time, potential impact range (such as affected equipment or loop number), and corresponding response suggestions. The structured format facilitates integration with SCADA platforms, master station dispatch systems, or mobile inspection terminals, improving the efficiency of early warning information dissemination and the human-machine interaction experience.

[0107] Meanwhile, in this embodiment, the generated early warning information is sent to the control system at the emergency power station site through the communication gateway. Based on different alarm levels, the following automated response strategies are executed: (1) For general alarms, the event log is recorded and a manual inspection reminder is triggered, requiring no immediate intervention; (2) For important alarms, the load management module is linked to perform zoned load limiting or scheduling adjustments, and a maintenance request is sent to the operation and maintenance system; (3) For emergency alarms, control commands such as circuit breaker tripping, backup power switching, and relay protection actions are immediately executed to ensure equipment and personnel safety. This response mechanism realizes a closed-loop process from "intelligent identification" to "rapid handling."

[0108] In step 700 of this embodiment, to achieve continuous optimization of model parameters and protection of local privacy data, a federated learning mechanism is used to periodically jointly train the time series prediction model and the Transformer state fusion network. Specifically, a fixed federated training cycle (such as daily or weekly) is set in the remote monitoring platform. At the beginning of each training round, the central coordinator distributes the initial parameters of the current global model to multiple distributed participating nodes, i.e., the edge computing units of each substation. Each node only receives the parameters and does not transmit local raw data, thereby ensuring that sensitive information on site does not leave the network side and meeting data compliance and security requirements.

[0109] In each local node, this embodiment utilizes stored historical monitoring data (including unified feature vector sequences and corresponding predicted labels) to train the local replica's time-series prediction model and Transformer state fusion network. Local training employs a fixed number of iterations or error convergence conditions, and the model parameter update is calculated after training. Each node uploads its local update results to the remote monitoring platform, where the central coordinator integrates the parameter update results using a weighted average or adaptive aggregation strategy to complete a unified update of the global model. This entire process achieves model knowledge sharing without aggregating any original data, improving global prediction accuracy and fault identification robustness.

[0110] Furthermore, to improve overall operational efficiency and edge-to-edge collaboration capabilities, this embodiment performs a centralized replay verification and performance evaluation after model updates. If the model meets the standards in key indicators such as prediction accuracy, fault identification accuracy, and false alarm rate, the new model is automatically activated and deployed to edge nodes. Simultaneously, based on the updated model's contribution evaluation results for each feature channel, the operating parameters of heterogeneous sensors are dynamically adjusted, including sampling frequency, data reporting cycle, and energy consumption control configuration. For example, low-contribution sensors will enter a low-power standby state, while high-value sensors will increase their sampling density, ensuring optimal allocation of global sensing capabilities under resource constraints, thereby constructing a sustainable and self-evolving intelligent monitoring system.

[0111] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides a remote intelligent operation monitoring system for intelligent substations, including:

[0112] The heterogeneous sensing and acquisition unit is used to deploy heterogeneous sensors on the primary equipment, auxiliary equipment and substation environment of the substation, so as to collect electrical parameters, environmental parameters and external meteorological information in real time through the heterogeneous sensors, and to align the electrical parameters, environmental parameters and external meteorological information of different protocols with timestamps to form a structured raw dataset.

[0113] The feature encoding processing unit is used to perform normalization processing on the numerical data in the structured original dataset to obtain a first processing result, and to perform word embedding encoding on the text data to obtain a second processing result, and to merge the first processing result and the second processing result into a unified feature vector sequence;

[0114] An encrypted transmission unit is used to encapsulate and encrypt the unified feature vector sequence and send it to a remote monitoring platform via a 5G mobile communication link.

[0115] The temporal modeling and prediction unit is used to perform multi-layer modeling on the unified feature vector sequence based on a temporal prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, and output the prediction sequence of key operating parameters within a future preset time window.

[0116] The state fusion evaluation unit is used to input the structured original dataset and the predicted sequence of key operating parameters into the Transformer state fusion network, generate a joint state vector, and output the probability of occurrence of each fault type and the corresponding confidence interval.

[0117] The early warning linkage control unit is used to automatically generate early warning information and simultaneously trigger control, maintenance and protection measures when the probability of any of the aforementioned fault types exceeds a set threshold.

[0118] The model adaptive optimization unit is used to update the parameters of the time-series prediction model and the Transformer state fusion network based on a federated learning mechanism using historical monitoring data at a preset period, and to dynamically manage the heterogeneous sensor nodes.

[0119] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0120] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A remote intelligent operation monitoring method for a smart substation, characterized in that, include: Heterogeneous sensors are deployed on primary equipment, auxiliary equipment, and the substation environment to collect electrical parameters, environmental parameters, and external meteorological information in real time. The electrical parameters, environmental parameters, and external meteorological information of different protocols are timestamped to form a structured raw dataset. Normalization is performed on the numerical data in the structured original dataset to obtain a first processing result, and word embedding encoding is performed on the text data to obtain a second processing result. The first processing result and the second processing result are then merged into a unified feature vector sequence. The unified feature vector sequence is encapsulated and encrypted and sent to the remote monitoring platform via a 5G mobile communication link; Based on the time-series prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, multi-layer modeling is performed on the unified feature vector sequence to output the prediction sequence of key operating parameters within a future preset time window. The structured original dataset and the predicted sequence of key operating parameters are input into the Transformer state fusion network to generate a joint state vector and output the probability of occurrence of each fault type and the corresponding confidence interval. When the probability of any of the aforementioned fault types exceeds a set threshold, an early warning message is automatically generated and control, maintenance, and protection measures are triggered simultaneously. The parameters of the time-series prediction model and the Transformer state fusion network are updated using historical monitoring data at preset intervals based on a federated learning mechanism, and the heterogeneous sensor nodes are dynamically managed.

2. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, Heterogeneous sensors are deployed on primary equipment, auxiliary equipment, and the substation environment to collect electrical parameters, environmental parameters, and external meteorological information in real time. The electrical parameters, environmental parameters, and external meteorological information from different protocols are timestamped to form a structured raw dataset, including: Establish a sensor deployment mapping table, plan the deployment location, measurement type and unique number of the heterogeneous sensors for primary equipment, auxiliary equipment and station environment of substation respectively, and record the physical coordinates of each heterogeneous sensor; The sampling tasks of all heterogeneous sensors are started synchronously through a unified broadcast trigger signal, and a local timestamp is generated for each frame of sampled data to form a synchronous sampling data frame. A time calibration mechanism based on a reference clock is adopted. For each heterogeneous sensor numbered k, within a preset calibration period, the time offset between the local timestamp sequence and the main reference clock sequence is calculated, and the original timestamp is corrected to an aligned timestamp. The formula for calculating the aligned timestamp is as follows: ;in, The original timestamp, representing the local location of sensor number k in the nth sampling; This represents the time offset between the sensor numbered k and the reference clock within the current calibration cycle, calculated as the median of multiple time differences. ; This represents the timestamp recorded by the reference clock during the i-th calibration broadcast; The aligned timestamp represents the sensor numbered k after the nth sampling. This indicates that within a complete calibration cycle, the reference clock and the numbered clock are... The number of times the sensor successfully established communication and recorded the timestamp difference; Based on a unified data encapsulation structure, the number, aligned timestamp, and corresponding measurement data triplet are combined. The original structured dataset is organized into a standardized format; wherein, For measurement data, it represents the physical quantity measurement value recorded by the sensor numbered k during the nth sampling; this measurement value is an electrical parameter, environmental parameter, or external meteorological information.

3. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, Normalization is performed on the numerical data in the structured original dataset to obtain a first processing result, and word embedding encoding is performed on the text data to obtain a second processing result. The first processing result and the second processing result are then merged into a unified feature vector sequence, including: For each numerical feature numbered m in the structured original dataset, calculate the 10th percentile of the numerical feature within a sliding window. 50th percentile 90th percentile , used for dynamic normalization processing; Using the normalization formula Numerical feature values ​​for each record in the structured dataset Normalization is performed to obtain the normalized result. ; For each text class feature tag in the record Using a pre-trained word embedding matrix One-hot encoded vector Projection as embedding vector ; Index the records as All normalized numerical features form a column vector All text is embedded into a column vector. And generate a unified feature vector through vector-level concatenation. Unified feature vector The calculation formula is as follows: ; in, This represents the original value of the numerical feature numbered m in the r-th global record of the structured original dataset; To avoid positive numbers that are divided by zero; This represents the normalization result of feature number m at record r; This is the one-hot encoded vector used to record the text features of class q in r; The embedding vector is used to record the corresponding text features in r.

4. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, The unified feature vector sequence is encapsulated and encrypted and sent to the remote monitoring platform via a 5G mobile communication link, including: Each feature vector in the unified feature vector sequence is segmented into basic data blocks according to a fixed byte length, and each data block is marked with an index number and a timestamp. Each data block is encrypted using the national standard SM4 block cipher algorithm with a unified key. and fixed initialization vector A ciphertext block sequence is generated; the formula for calculating the ciphertext block sequence is: ;in, Indicates the index in the unified feature vector sequence. eigenvectors; Indicates the use of a key and initialization vector For vectors Perform SM4 symmetric block encryption operation; For the corresponding ciphertext block sequence; The encrypted ciphertext block sequence is encapsulated into a data packet frame structure; the data packet frame structure includes a frame header, a frame body, and a frame trailer; the frame header includes the frame type, the number of data blocks, and a synchronization flag; the frame body is the ciphertext block content; the frame trailer is a CRC checksum; the CRC checksum is used for message integrity verification. By utilizing a 5G IoT module to access the public network, a remote monitoring platform address and port based on a multi-access edge computing mode are configured. The data packet frame structure is transmitted in an orderly manner through the TCP / IP protocol, and the transmission confirmation response is recorded and resuming interrupted transmission is supported.

5. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, Based on a time-series prediction model composed of a cascaded one-dimensional convolutional layer and a two-layer long short-term memory neural network, multi-layer modeling is performed on the unified feature vector sequence to output a predicted sequence of key operating parameters within a future preset time window, including: A fixed-length sliding window is used to slice the uniform feature vector sequence to obtain the input tensor. Each time step contains all feature dimensions; Use a length of One-dimensional convolution kernel sequence Perform convolution operations and obtain the convolution feature sequence through a linear rectified activation function; The convolutional feature sequences are sequentially input into the first layer of the Long Short-Term Memory (LSTM) network. 1， Extract short-term and long-term time dependencies; The hidden state sequence of LSTM1 is input into the second layer of the Long Short-Term Memory (LSTM) network. 2， Generate deep temporal representations; A linear transformation is performed on the current hidden state of the LSTM2, and the prediction vector is output through the Sigmoid activation function, as shown in the following formula: in, To unify the eigenvector sequence at time 1 The input tensor at the location; The first one in the one-dimensional convolution kernel sequence Each convolutional weight matrix; is the kernel length; ReLU is the linear rectified activation function; LSTM1 and LSTM2 represent the first and second layers of the long short-term memory network operations, respectively. To output the mapping matrix; This is the output bias vector; Use the Sigmoid activation function; Predicting step size relative to the current time The key operating parameter prediction vector.

6. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, The structured original dataset and the predicted sequence of key operating parameters are input into the Transformer state fusion network to generate a joint state vector, and the probability of occurrence of each fault type and the corresponding confidence interval are output, including: Under the same time index r, normalized numerical-text feature vectors in the original structured dataset With corresponding key operating parameter prediction vectors Perform vector-level concatenation to obtain the concatenated input vector. To concatenate the input vector sequence Superimposed learnable position embedding matrices To form an encoding matrix ; Encoding matrix The signals are sequentially fed into an L-layer multi-head self-attention Transformer encoder. The self-attention output is calculated within each layer using the following formula: Take the first row vector of the output matrix of the last encoder layer As a joint state vector, the fault probability vector is obtained by combining the joint state vector through a linear mapping-Softmax method. softmax ; During the inference phase, M random deactivated forward propagations are performed on the same input to obtain a probability sample set. The confidence interval for each fault type f is estimated using the mean-variance method, and the formula is as follows: ； ； ； in, For all A matrix arranged in chronological order; For the position embedding matrix; A query, key, and value mapping matrix; Hide dimensions in the model; To map weights and biases for classification; This is the failure probability vector obtained from a single inference; Let f be the probability of fault type f in the m-th random inactivation inference; and These are the mean and variance of the probability of fault type f, respectively; With a 95% confidence level; The 95% confidence interval for fault type f; The length of the concatenated input vector sequence; M represents the number of samples for forward propagation performed using a random deactivation mechanism during the inference phase.

7. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, When the probability of any of the aforementioned fault types exceeds a set threshold, an early warning message is automatically generated, and control, maintenance, and protection measures are triggered simultaneously, including: Establish a mapping table between fault types and response levels, set response thresholds for various fault types, and classify them into three levels according to severity: general alarms, important alarms, and emergency alarms. The probability of each type of fault output by the state fusion network is compared item by item. If any fault probability is higher than its corresponding alarm threshold, it is determined to be a fault warning trigger event. Based on the matched alarm level, structured early warning information is generated in the monitoring platform; the early warning information includes the fault type, probability of occurrence, estimated time, scope of impact, and response suggestions. The warning information is synchronously sent to the local substation control device through the communication gateway, and the corresponding response strategy is executed according to the alarm level.

8. The remote intelligent operation monitoring method for intelligent substations according to claim 7, characterized in that, The response strategy includes: General alarm trigger status records and manual inspection prompts; Critical alarms trigger partition load adjustments and planned maintenance requests; Emergency alarms directly trigger fault isolation, backup power switching, and protection actions.

9. The remote intelligent operation monitoring method for intelligent substations according to claim 1, characterized in that, Using historical monitoring data at preset intervals, the parameters of the time-series prediction model and the Transformer state fusion network are updated based on a federated learning mechanism, and the heterogeneous sensor nodes are dynamically managed, including: A fixed federated training cycle is set in the remote monitoring platform. At the beginning of each cycle, the initial parameters of the model are sent from the central coordinator to each participating site. Historical monitoring data stored in each substation's local node is retrieved and used to train the time-series prediction model and the local copy of the Transformer state fusion network, respectively, and the local model update results are recorded. The parameter updates of the local model are sent back to the remote monitoring platform, where the central coordinating unit performs parameter aggregation without aggregating the original data, thus completing the global model update. The updated model is replayed for verification and performance evaluation. If the evaluation results meet the target conditions of improved accuracy and reduced false alarm rate, the new model is activated for subsequent prediction and fusion tasks. After the model update is completed, the sampling frequency, reporting period and power consumption configuration of the heterogeneous sensor are dynamically adjusted based on the feature redundancy and sampling contribution feedback from the model, so as to achieve adaptive optimization management of the nodes of the heterogeneous sensor.

10. A remote intelligent operation monitoring system for an intelligent substation, characterized in that, include: The heterogeneous sensing and acquisition unit is used to deploy heterogeneous sensors on the primary equipment, auxiliary equipment and substation environment of the substation, so as to collect electrical parameters, environmental parameters and external meteorological information in real time through the heterogeneous sensors, and to align the electrical parameters, environmental parameters and external meteorological information of different protocols with timestamps to form a structured raw dataset. The feature encoding processing unit is used to perform normalization processing on the numerical data in the structured original dataset to obtain a first processing result, and to perform word embedding encoding on the text data to obtain a second processing result, and to merge the first processing result and the second processing result into a unified feature vector sequence; An encrypted transmission unit is used to encapsulate and encrypt the unified feature vector sequence and send it to a remote monitoring platform via a 5G mobile communication link. The temporal modeling and prediction unit is used to perform multi-layer modeling on the unified feature vector sequence based on a temporal prediction model composed of a one-dimensional convolutional layer and a two-layer long short-term memory neural network, and output the prediction sequence of key operating parameters within a future preset time window. The state fusion evaluation unit is used to input the structured original dataset and the predicted sequence of key operating parameters into the Transformer state fusion network, generate a joint state vector, and output the probability of occurrence of each fault type and the corresponding confidence interval. The early warning linkage control unit is used to automatically generate early warning information and simultaneously trigger control, maintenance and protection measures when the probability of any of the aforementioned fault types exceeds a set threshold. The model adaptive optimization unit is used to update the parameters of the time-series prediction model and the Transformer state fusion network based on a federated learning mechanism using historical monitoring data at a preset period, and to dynamically manage the heterogeneous sensor nodes.

Images