Transformer area situation awareness method and system based on intelligent analysis terminal

Abstract

The invention discloses a transformer area situation awareness method and system based on an intelligent analysis terminal, and relates to the technical field of transformer area power monitoring. The method comprises the following steps: deploying an intelligent analysis terminal in a transformer in a transformer area, and synchronously collecting electrical data, infrastructure monitoring data and meteorological data of the transformer area for preprocessing; the intelligent analysis terminal carries out time scale synchronous alignment on the preprocessed data; performing cross-modal data fusion on the data after time scale synchronization; constructing an LSTM network to predict communication time delay; processing the abnormal data; and constructing a performance evaluation index to evaluate the training model. The method comprises the following steps: deploying an intelligent analysis terminal in a transformer area transformer, carrying out time scale synchronous alignment on data, and constructing an LSTM network to predict communication time delay; a performance evaluation index is constructed to evaluate a training model, a digital twin test environment is constructed, asynchronous data is injected to verify the fault-tolerant capability of the system, and the monitoring accuracy and the data resource utilization rate of a transformer area are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power monitoring in a distribution area, and particularly relates to a method and system for distribution area situation awareness based on an intelligent analysis terminal. Background Art

[0002] With the increasing scale of the distribution network, the operation of the distribution network tends to be more complex and diverse, and the requirements for operation management are getting higher and higher. The management of the distribution network includes the management of low-voltage distribution transformers, and the stable operation of its distribution transformers is an important prerequisite for ensuring the safety and reliability of the distribution network.

[0003] The problems faced by the operation state monitoring of distribution transformers in the distribution area are mainly reflected in the following two aspects:

[0004] (1) The state perception types of the monitoring system for distribution transformers in the distribution area are single and the dimensions are incomplete. The initial data for the state evaluation of distribution transformers in the distribution area are not comprehensive, which hinders the subsequent accurate monitoring of the state of distribution transformers in the distribution area;

[0005] (2) The problem of information islands among the monitoring systems for distribution transformers in the distribution area is prominent, and the utilization efficiency of distribution transformer data resources is not high; the functions are single, and the systems are relatively isolated from each other without information sharing, resulting in the inability to perform panoramic situation awareness on distribution transformers in the distribution area. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for distribution area situation awareness based on an intelligent analysis terminal. By deploying an intelligent analysis terminal on a distribution transformer in the distribution area, synchronizing and aligning data on a time scale, constructing an LSTM network to predict communication delay; constructing performance evaluation indicators to evaluate the training model, and constructing a digital twin test environment to inject asynchronous data to verify the fault tolerance ability of the system, the problems of inaccurate existing distribution area monitoring, low utilization efficiency of data resources, and inaccurate situation awareness are solved.

[0007] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0008] The present invention provides a method for distribution area situation awareness based on an intelligent analysis terminal, including the following steps:

[0009] Step S1: Deploy an intelligent analysis terminal on a distribution transformer in the distribution area, synchronously collect electrical data, infrastructure monitoring data, and meteorological data in the distribution area for preprocessing;

[0010] Step S2: The intelligent analysis terminal synchronizes and aligns the preprocessed data on a time scale;

[0011] Step S3: Perform cross-modal data fusion on the data synchronized on the time scale;

[0012] Step S4: Construct an LSTM network to predict communication delay;

[0013] Step S5: Process abnormal data;

[0014] Step S6: Construct performance evaluation metrics to evaluate the trained model, and construct a digital twin test environment to inject asynchronous data to verify the fault tolerance ability of the system.

[0015] As a preferred technical solution, in the step S1, the monitoring master station broadcasts a time synchronization instruction to the intelligent analysis terminal through the SNPT protocol to ensure that the clock error of all nodes is < 10 ms. The monitoring master station realizes the time calibration of the intelligent analysis terminal (accuracy ±1 μs) by deploying a Beidou / GPS dual-mode time service module;

[0016] The structured data collected by the intelligent analysis terminal is encapsulated according to the IEC61850 standard. The encapsulation includes five-dimensional time tags, such as acquisition time, transmission delay, data validity period, clock source status, and check code, to realize second-level data validity verification; the unstructured data (such as video, image, audio) collected by the intelligent analysis terminal is compressed by H.265 encoding. Each key frame is attached with the device GPS coordinates and timestamp, and the audio signal is encoded by G.711 and superimposed with a 10 kHz synchronous pilot signal for delay calibration.

[0017] As a preferred technical solution, in the step S2, when aligning structured data, a sliding time window is constructed, and cubic spline interpolation is used to complement missing data; when aligning unstructured data, the video stream and SCADA data are synchronized through time axis mapping.

[0018] As a preferred technical solution, in the step S3, a spatio-temporal correlation matrix is established:

[0019]

[0020] where M i,j represents the spatio-temporal correlation weight between data points i and j, with a range of [0,1], t i , t j respectively represent the timestamps of data points i and j, loc i , loc jRepresent the spatial coordinates of data points i and j respectively, α and β represent the time attenuation coefficient and the spatial attenuation coefficient respectively, and ||·|| represents the spatial distance. By integrating heterogeneous data such as sensors, videos, and logs, we can break through the cognitive limitations of a single data source and establish a global view under a unified spatiotemporal benchmark. Through spatiotemporal constraints, we can mine the potential associations of cross-modal data (such as the linkage changes between equipment failures and surrounding video surveillance images) and capture implicit rules that are difficult to discover with traditional methods. The spatiotemporal association matrix assigns dynamic weights to different data sources (such as giving priority to trusting data from spatiotemporal neighboring nodes), suppress noise interference, and improve prediction robustness.

[0021] As a preferred technical solution, in step S4, the specific steps of constructing an LSTM network to predict communication delay are as follows:

[0022] Step S41: collect spatial coding, time coding and fusion strategy for spatiotemporal alignment; spatial coding converts the device GPS coordinates into 32-dimensional space vectors through the H3 geographic grid system to capture the topological relationship of device distribution; time coding decomposes the timestamp at multiple scales (seconds, minutes, hours) and superimposes periodic coding (working days / holidays, seasonal factors); according to the previous time scale alignment method, ensure that the timestamp error of multi-source data is less than 20ms and the spatial coordinate error is less than 10 meters;

[0023] The fusion strategy uses a fully connected network to combine the spatiotemporal vector with the device status (CPU load, signal strength) and network indicators (bandwidth utilization, packet loss rate) into 128-dimensional joint features. Compared with the traditional LSTM that only considers the time dimension, this solution quantifies the spatial distance into a topological relationship through H3 geocoding, so that the model can identify the spatial correlation pattern of the device group (such as the chain delay effect of devices in the power grid area).

[0024] Step S42: Automatically adjust the input sequence length (default 60 minutes) according to business needs, and use overlapping sampling strategy to improve data utilization;

[0025] Step S43: training the benchmark model on the historical data set, freezing the underlying LSTM parameters, and generating a universal spatiotemporal representation;

[0026] Step S44: After being deployed to the edge node, the top three layer parameters are opened for online fine-tuning, and the learning rate is set to 1 / 10 of that in the pre-training stage; the federated incremental learning mechanism enables the model to maintain global consistency while quickly adapting to local network changes while protecting data privacy;

[0027] Step S45: injecting Gaussian noise, random packet loss, etc. to perform adversarial enhancement training to improve the robustness of the model to network fluctuations;

[0028] Step S46: remove the connections with weights less than 0.1 in the attention layer and prune the computation graph;

[0029] Step S47: Perform dynamic precision switching. Use FP16 inference in the normal state and automatically switch to FP32 precision when a decrease in prediction confidence is detected. At the same time, deploy model instances with both FP16 and FP32 precisions, share weight parameters through memory mapping technology to reduce storage overhead, establish a precision switching channel, and maintain a shared intermediate calculation result buffer in the GPU video memory to achieve zero-copy data transfer during precision mode switching.

[0030] When performing seamless switching, use a hardware-accelerated precision conversion unit: NVIDIA Tensor Core automatically processes the format alignment of FP16 to FP32. During the switching process, automatically reorganize low-confidence samples into independent micro-batches (1-4 samples) to avoid delays caused by global precision conversion.

[0031] When performing adaptive threshold adjustment, dynamically adjust the confidence threshold based on historical data distribution. The specific calculation formula is as follows:

[0032]

[0033] In the formula, θ base is the reference threshold (default is 0.7), and σ is the standard deviation of the metric within the past 1 hour.

[0034] Step S48: Use Kalman filtering to smooth the output delay sequence of the network prediction results, and at the same time output the delay mean, fluctuation range, and anomaly probability to meet different business requirements: the mean is used for resource scheduling, the range is used for risk assessment, and the anomaly probability is used for warning threshold setting.

[0035] As a preferred technical solution, in step S46, introduce a dynamic gating mechanism in the multi-head attention layer, automatically close the connection channels with an absolute weight value <0.1 through learnable gating parameters, use Taylor expansion importance evaluation to calculate the influence degree of each attention head on the loss function, and remove the attention heads with an influence degree lower than the threshold (such as 0.05).

[0036] When performing neuron-level pruning on the LSTM unit, based on the L2 norm sorting of the hidden state activation values, delete 20% of the neurons with the lowest long-term activation intensity. When performing filter-level pruning on the federated learning node, count the gradient magnitudes of the edge device local data on the LSTM weights and remove the filters with a gradient mean <1e-5.

[0037] When initializing the global model, apply L1 sparse regularization to the LSTM layer and the attention layer, force more than 50% of the connection weights to approach zero, and use iterative progressive pruning. After each 10% pruning operation, perform 2 rounds of fine-tuning until the model parameter quantity is reduced by 40% and the accuracy loss <3%.

[0038] When performing local training on edge nodes, the pruning threshold is dynamically adjusted based on the moving average importance score. The specific formula is: importance score = 0.7 * absolute value of weight + 0.3 * variance of gradient. When the central server aggregates, the mask alignment technology is adopted: only the non-zero parameter positions shared by all edge nodes are retained, eliminating the differences in pruning structures between devices.

[0039] The pruned zero weights are converted into computational graph jump instructions, reducing 60% of the ineffective calculations of matrix multiplication, generating a sparse computational graph. According to the complexity of the spatio-temporal features of the input data, 30% - 80% of the LSTM neurons are automatically activated to achieve energy consumption-sensitive inference.

[0040] As a preferred technical solution, in step S6, the constructed performance evaluation indicators include the time alignment rate indicator and the maximum absolute value indicator; when constructing the digital twin test environment, a distributed sensor network is deployed to cover the key nodes of the physical system, and millisecond-level synchronous acquisition is achieved through the 5G / TSN protocol. Unstructured data sources are integrated, and the H.265 encoding is used to compress the transmission bandwidth to ensure that the spatio-temporal alignment error of multi-modal data is < 50 ms; at the same time, based on the CAD model of the physical system, a three-dimensional digital model is reconstructed through point cloud scanning and deep learning to achieve sub-millimeter geometric accuracy; a multi-scale simulation engine is embedded to support the switching of test scenarios with different granularities.

[0041] As a preferred technical solution, in step S6, the constructed performance evaluation indicators include the time alignment rate indicator and the maximum absolute value indicator; when constructing the digital twin test environment, a distributed sensor network is deployed to cover the key nodes (such as device status, environmental parameters) of the physical system, and millisecond-level synchronous acquisition is achieved through the 5G / TSN protocol. Unstructured data sources (video streams, voice instructions) are integrated, and the H.265 encoding is used to compress the transmission bandwidth to ensure that the spatio-temporal alignment error of multi-modal data is < 50 ms; at the same time, based on the CAD model of the physical system, a three-dimensional digital model is reconstructed through point cloud scanning and deep learning to achieve sub-millimeter geometric accuracy; a multi-scale simulation engine is embedded to support the switching of test scenarios with different granularities.

[0042] The present invention is a substation area situation awareness system based on an intelligent analysis terminal, including a security and resilience data fusion unit and a security and resilience perception and warning unit;

[0043] The safety and resilience data fusion unit includes a multi-source pre-purchase data acquisition module, a time scale alignment module, a feature extraction module, a safety assessment and analysis module, and a fault prediction module; the multi-source pre-purchase data acquisition module is used to collect meteorological data, electrical data, and infrastructure monitoring data; the time scale alignment module is used to align the time scales of structured data and unstructured data; the feature extraction module is used to extract features from the data with aligned time scales; the safety assessment and analysis module is used to evaluate the trained model; the fault prediction module is used to predict faults in the distribution transformer area.

[0044] The safety and resilience perception and early warning unit includes network terminal devices, a big data analysis module, a real-time monitoring module, a deep learning module, an intelligent decision support system, an optimization scheduling module, a fault analysis module, and an intelligent repair module; the network terminal devices are used to collect distribution transformer data; the real-time monitoring module is used to conduct real-time monitoring of the distribution transformer area; the deep learning module is used to predict communication delays through an LSTM network; the intelligent decision support system is used to provide decision-making assistance during the occurrence of distribution transformer accidents; the optimization scheduling module is used to allocate personnel and resources after an accident occurs; the fault analysis module is used to analyze the faults of the distribution transformer after a fault occurs; the intelligent repair module is used to generate a repair plan for the distribution transformer after a fault occurs.

[0045] As a preferred technical solution, before an accident occurs, the historical operation data and environmental parameters of the distribution transformer are mined by comprehensively applying big data analysis technology and machine learning algorithms to predict potential fault hazards; during extreme event conditions, manual equipment inspections are replaced to collect the status data of the equipment in real time; during the accident process, real-time monitoring and emergency response are carried out, deep learning and reinforcement learning are introduced to enhance the capabilities of the SCADA system, and real-time analysis and prediction of the status of the distribution transformer are carried out; the intelligent decision support system provides the optimal emergency response plan; in the later stage of the accident, big data analysis technology and AI algorithms are used to carry out fault analysis and intelligent repair, identify fault patterns and root causes, automatically adjust the operation parameters of the distribution transformer, and perform remote self-repair through the Internet of Things.

[0046] The present invention has the following beneficial effects:

[0047] The present invention deploys an intelligent analysis terminal on the distribution transformer, synchronizes and aligns the time scales of the data, constructs an LSTM network to predict communication delays; constructs performance evaluation indicators to evaluate the trained model, and constructs a digital twin test environment to inject asynchronous data to verify the fault tolerance ability of the system, improving the accuracy of distribution transformer monitoring and the utilization rate of data resources.

[0048] Through cross-modal data fusion, the present invention establishes a spatio-temporal correlation matrix, integrates heterogeneous data such as sensors, videos, and logs, breaks through the cognitive limitations of a single data source, establishes a global view under a unified spatio-temporal benchmark, and mines potential correlations in cross-modal data through spatio-temporal constraints, captures hidden rules that are difficult to discover by traditional methods, eliminates information silos, enhances pattern recognition, and improves prediction robustness.

[0049] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for describing the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 It is a flowchart of a method for detecting the situation of a power distribution area based on an intelligent analysis terminal according to the present invention;

[0052] Figure 2 It is a schematic structural diagram of a system for detecting the situation of a power distribution area based on an intelligent analysis terminal according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0054] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0055] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further describe the present application in detail with reference to Figure 1-2 the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0056] Embodiment 1

[0057] Please refer to Figure 1 As shown, the present invention is a method for detecting the situation of a power distribution area based on an intelligent analysis terminal, including the following steps:

[0058] Step S1: Deploy an intelligent analysis terminal at the substation transformer, synchronously collect electrical data, infrastructure monitoring data, and meteorological data of the substation area, and perform preprocessing;

[0059] Step S2: The intelligent analysis terminal synchronizes and aligns the preprocessed data in terms of time scale;

[0060] Step S3: Perform cross-modal data fusion on the data after time-scale synchronization;

[0061] Step S4: Construct an LSTM network to predict communication delay;

[0062] Step S5: Process abnormal data;

[0063] Step S6: Construct performance evaluation indicators to evaluate the training model, and construct a digital twin test environment to inject asynchronous data to verify the fault tolerance ability of the system.

[0064] In Step S1, the monitoring master station broadcasts a time calibration instruction to the intelligent analysis terminal through the SNPT protocol to ensure that the clock error of all nodes is < 10 ms. The monitoring master station realizes the time calibration of the intelligent analysis terminal (accuracy ±1 μs) by deploying a Beidou / GPS dual-mode time service module;

[0065] Structured data refers to data types with a predefined data model and strict format planning, which can be logically expressed through a two-dimensional table structure (rows + columns), such as power data (voltage value, current phase, etc.), equipment status data (circuit breaker opening and closing status, working temperature, etc.), and meteorological numerical data (wind speed, temperature, etc.);

[0066] Unstructured data refers to information types without a fixed format or pattern, including various data forms, which require special processing to extract value; such as text, images, audio, video, spectrum, etc.

[0067] The structured data collected by the intelligent analysis terminal is encapsulated according to the IEC61850 standard. The encapsulation includes five-dimensional time tags, such as acquisition time, transmission delay, data validity period, clock source status, and check code, to achieve second-level data validity verification; the unstructured data (such as video, image, audio) collected by the intelligent analysis terminal is compressed by H.265 encoding, and each key frame is attached with the device GPS coordinates and time stamp, while the audio signal is encoded by G.711 and superimposed with a 10 kHz synchronization pilot signal for delay calibration.

[0068] In Step S3, establish a spatio-temporal correlation matrix:

[0069]

[0070] In the formula, M i,j represents the spatio-temporal correlation weight between data points i and j, with a range of [0,1], ti ,t j Respectively represent the timestamps of data points i and j, loc i ,loc j They represent the spatial coordinates of data points i and j respectively, α and β represent the time attenuation coefficient and the spatial attenuation coefficient respectively, and ||·|| represents the spatial distance.

[0071] Temporal correlation is modeled by the time difference |t i -t j | and spatial distance ||loc i -loc j || linear combination, quantifying the spatiotemporal correlation between multi-source data. Weight value M i,j It decays with the increase of time and space differences and is used in scenarios such as data fusion and anomaly detection.

[0072] The parameters of α and β can be adjusted dynamically to suit different sensitive scenarios, such as:

[0073] Highly time-sensitive scenarios: Increase α, such as α = 0.1 / ms, to make the time difference more significant.

[0074] High spatial sensitivity scene: Increase β, such as β = 1.0 / m, the impact of spatial distance is more significant.

[0075] By integrating heterogeneous data such as sensors, videos, and logs, we can break through the cognitive limitations of a single data source and establish a global view under a unified spatiotemporal benchmark. Through spatiotemporal constraints, we can explore the potential correlations between cross-modal data (such as the linkage changes between equipment failures and surrounding video surveillance images) and capture implicit rules that are difficult to discover with traditional methods. The spatiotemporal correlation matrix gives dynamic weights to different data sources (such as giving priority to trusting data from spatiotemporal neighboring nodes), suppressing noise interference and improving prediction robustness.

[0076] For example, when conducting situation awareness in a substation, the data is integrated: meter voltage data (time series) + infrared camera temperature data (image) + drone inspection coordinates (space); A spatiotemporal correlation matrix is established, where α = 0.05 / min (time attenuation coefficient) and β = 0.8 / km (spatial attenuation coefficient); therefore, when a situational awareness failure occurs in the substation, the matrix automatically highlights the associated equipment (correlation coefficient > 0.85), and the positioning time is shortened from 2 hours to 15 minutes.

[0077] In step S4, the specific steps of constructing an LSTM network to predict communication delay are as follows:

[0078] Step S41: Collect spatial encoding, temporal encoding, and fusion strategy for spatio-temporal alignment; The spatial encoding converts the device GPS coordinates into 32-dimensional spatial vectors through the H3 geospatial grid system to capture the topological relationship of device distribution; The temporal encoding performs multi-scale decomposition (seconds, minutes, hours) on timestamps and superimposes periodic encoding (weekday / holiday, seasonal factor); According to the previous time scale alignment method, ensure that the timestamp error of multi-source data is <20ms and the spatial coordinate error is <10 meters;

[0079] The fusion strategy concatenates spatio-temporal vectors with device status (CPU load, signal strength) and network metrics (bandwidth utilization, packet loss rate) into 128-dimensional joint features through a fully connected network; Compared with traditional LSTM that only considers the time dimension, this solution quantifies the spatial distance into a topological relationship through H3 geocoding, enabling the model to identify the spatial association patterns of device groups (such as the cascading delay effect of devices within a power grid substation area);

[0080] Step S42: Automatically adjust the input sequence length according to business requirements (default 60 minutes), and adopt an overlapping sampling strategy to improve data utilization;

[0081] Step S43: Train a baseline model on the historical dataset, freeze the underlying LSTM parameters, and generate a general spatio-temporal representation;

[0082] After deploying to edge nodes, open the top 3-layer parameters for online fine-tuning, and set the learning rate to 1 / 10 of the pre-training stage; The federated incremental learning mechanism enables the model to maintain global consistency and quickly adapt to local network changes while protecting data privacy;

[0083] Step S45: Inject Gaussian noise, random packet loss, etc. for adversarial enhancement training to improve the robustness of the model to network fluctuations;

[0084] Step S46: Remove the connections with weights less than 0.1 in the attention layer and perform computational graph pruning;

[0085] Perform dynamic precision switching, use FP16 inference in normal state, and automatically switch to FP32 precision when the predicted confidence level drops; At the same time, deploy model instances of both FP16 and FP32 precisions, share weight parameters through memory mapping technology to reduce storage overhead; Establish a precision switching channel, maintain a shared intermediate calculation result buffer in the GPU video memory, and achieve zero-copy data transfer during precision mode switching;

[0086] When performing seamless switching, a hardware-accelerated precision conversion unit is adopted: NVIDIA Tensor Core automatically processes the format alignment of FP16 to FP32; during the switching process, low-confidence samples are automatically recombined into independent micro-batches (1-4 samples) to avoid the latency caused by global precision conversion.

[0087] When performing adaptive threshold adjustment, the confidence threshold is dynamically adjusted based on the historical data distribution. The specific calculation formula is as follows:

[0088]

[0089] In the formula, θ base is the reference threshold (default is 0.7), and σ is the standard deviation of the indicator within the past 1 hour.

[0090] Step S48: Use Kalman filtering to smooth the output delay sequence of the network prediction results, and at the same time output the delay mean, fluctuation range, and anomaly probability to meet different service requirements: the mean is used for resource scheduling, the range is used for risk assessment, and the anomaly probability is used for warning threshold setting;

[0091] Step S48: Use Kalman filtering to smooth the output delay sequence of the network prediction results, and at the same time output the delay mean, fluctuation range, and anomaly probability to meet different service requirements: the mean is used for resource scheduling, the range is used for risk assessment, and the anomaly probability is used for warning threshold setting.

[0092] In step S46, a dynamic gating mechanism is introduced in the multi-head attention layer. The connection channels with absolute weight <0.1 are automatically closed through learnable gating parameters. The importance evaluation of Taylor expansion is adopted to calculate the influence degree of each attention head on the loss function, and the attention heads with influence degree lower than the threshold (such as 0.05) are removed;

[0093] When performing neuron-level pruning on the LSTM unit, based on the L2 norm sorting of the hidden state activation values, the 20% neurons with the lowest long-term activation intensity are deleted. When performing filter-level pruning on the federated learning node, the gradient amplitude of the local data of the edge device on the LSTM weight is statistically calculated, and the filters with gradient mean <1e-5 are removed.

[0094] When initializing the global model, L1 sparse regularization is applied to the LSTM layer and the attention layer, forcing more than 50% of the connection weights to approach zero, and iterative progressive pruning is adopted. After each 10% pruning operation, 2 rounds of fine-tuning are performed until the model parameter quantity is reduced by 40% and the accuracy loss <3%.

[0095] When performing local training on edge nodes, the pruning threshold is dynamically adjusted based on the moving-average importance score. The specific formula is: Importance score = 0.7 * absolute value of weight + 0.3 * variance of gradient. When aggregating at the central server, the mask alignment technique is adopted: only the non-zero parameter positions shared by all edge nodes are retained, eliminating the pruning structure differences between devices.

[0096] The pruned zero weights are converted into computational graph jump instructions, reducing the ineffective calculations of 60% matrix multiplications, generating a sparse computational graph. According to the spatio-temporal feature complexity of the input data, 30% - 80% of the LSTM neurons are automatically activated to achieve energy consumption-sensitive inference.

[0097] As a preferred technical solution, in step S5, the three-level anomaly determination rules are defined as follows in the table:

[0098]

[0099]

[0100] In step S6, the constructed performance evaluation metrics include the time alignment rate metric and the maximum absolute value metric; when constructing the digital twin test environment, a distributed sensor network is deployed to cover the key nodes of the physical system (such as device status, environmental parameters), achieving millisecond-level synchronous acquisition through the 5G / TSN protocol, integrating unstructured data sources (video streams, voice instructions), using H.265 encoding to compress the transmission bandwidth, ensuring that the spatio-temporal alignment error of multi-modal data is < 50 ms; at the same time, based on the CAD model of the physical system, a three-dimensional digital model is reconstructed through point cloud scanning and deep learning, achieving sub-millimeter geometric accuracy; embedding a multi-scale simulation engine to support the switching of test scenarios with different granularities.

[0101] Embodiment 2

[0102] Refer to Figure 2 As shown, the present invention is a method and system for substation area situation awareness based on an intelligent analysis terminal, which can be used to execute the method content of Embodiment 1 of the present invention, including: a security resilience data fusion unit and a security resilience perception and warning unit;

[0103] The security resilience data fusion unit includes a multi-source pre-purchase data collection module, a time scale alignment module, a feature extraction module, a security assessment and analysis module, and a fault prediction module; the multi-source pre-purchase data collection module is used to collect meteorological data, electrical data, and infrastructure monitoring data; the time scale alignment module is used to perform time scale alignment on structured data and unstructured data; the feature extraction module is used to extract features from the time scale-aligned data; the security assessment and analysis module is used to evaluate the trained model; the fault prediction module is used to predict faults of the substation area transformer;

[0104] The safety and resilience perception and early warning unit includes network terminal devices, a big data analysis module, a real-time monitoring module, a deep learning module, an intelligent decision support system, an optimization scheduling module, a fault analysis module, and an intelligent repair module; the network terminal devices are used to collect data of the substation area transformers; the real-time monitoring module is used to conduct real-time monitoring of the substation area transformer area; the deep learning module is used to predict communication delay through the LSTM network; the intelligent decision support system is used to provide decision-making assistance during the occurrence of substation area accidents; the optimization scheduling module is used to allocate personnel and resources after the accident occurs; the fault analysis module is used to conduct fault analysis on the substation area transformers after the fault occurs; the intelligent repair module is used to generate a repair plan for the substation area transformers after the fault occurs.

[0105] As a preferred technical solution, before the accident occurs, comprehensively utilize big data analysis technology and machine learning algorithms to mine the historical operation data and environmental parameters of the substation area transformers, and predict potential fault hazards; replace manual equipment inspection under extreme event conditions and collect the status data of the equipment in real time; during the accident, conduct real-time monitoring and emergency response, introduce deep learning and reinforcement learning to enhance the capabilities of the SCADA system, and conduct real-time analysis and prediction of the status of the substation area transformers; the intelligent decision support system provides the optimal emergency response plan; in the later stage of the accident, utilize big data analysis technology and AI algorithms to conduct fault analysis and intelligent repair, identify the fault mode and root cause, automatically adjust the operation parameters of the substation area transformers, and conduct remote self-repair through the Internet of Things.

[0106] It should be noted that in the above system embodiments, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only for the convenience of mutual distinction and do not limit the protection scope of the present invention.

[0107] In addition, those of ordinary skill in the art can understand that all or part of the steps in implementing the methods of the above embodiments can be completed by instructing relevant hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0108] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. This specification selects and specifically describes these embodiments to better explain the principle and practical application of the present invention, so that those skilled in the art in the relevant technical field can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A method for sensing the situation of a substation area based on an intelligent analysis terminal, characterized in that: It includes the following steps: Step S1: Deploy an intelligent analysis terminal at the substation transformer, synchronously collect the electrical data, infrastructure monitoring data, and meteorological data of the substation area for preprocessing; Step S2: The intelligent analysis terminal synchronizes and aligns the preprocessed data in terms of time scale; Step S3: Perform cross-modal data fusion on the data after time-scale synchronization; Step S4: Construct an LSTM network to predict communication delay; Step S5: Process abnormal data; Step S6: Construct performance evaluation metrics to evaluate the training model, and construct a digital twin test environment to inject asynchronous data to verify the fault tolerance of the system.

2. The method for substation area situation awareness based on an intelligent analysis terminal according to claim 1, wherein In step S1, the monitoring master station broadcasts a time calibration instruction to the intelligent analysis terminal through the SNPT protocol, and the monitoring master station realizes the time calibration of the intelligent analysis terminal by deploying a Beidou / GPS dual-mode timing module; the structured data collected by the intelligent analysis terminal is encapsulated; the unstructured data collected by the intelligent analysis terminal is encoded and compressed, and each key frame is attached with the device GPS coordinates and timestamps.

3. A method for substation area situation awareness based on an intelligent analysis terminal according to claim 1, characterized in that, In step S2, when aligning structured data, a sliding time window is constructed, and cubic spline interpolation is used to complement missing data; when aligning unstructured data, the video stream and SCADA data are synchronized through time axis mapping.

4. A method for substation area situation awareness based on an intelligent analysis terminal according to claim 1, characterized in that In step S3, a spatio-temporal correlation matrix is established: where M i,j represents the spatio-temporal correlation weight between data points i and j, with a range of [0, 1], and t i , t j represent the timestamps of data points i and j respectively, and loc i , loc j represent the spatial coordinates of data points i and j respectively. α and β represent the time decay coefficient and the spatial decay coefficient respectively, and ||·|| represents the spatial distance.

5. A method for substation area situation awareness based on an intelligent analysis terminal according to claim 1, characterized in that, In step S4, the specific steps for constructing an LSTM network to predict communication delay are as follows: Step S41: Collect spatial encoding, time encoding, and fusion strategies for spatio-temporal alignment; Step S42: Automatically adjust the input sequence length according to service requirements; Step S43: Train a baseline model on the historical dataset, freeze the underlying LSTM parameters, and generate a general spatio-temporal representation; Step S44: After deploying to the edge node, open the parameters of the top 3 layers for online fine-tuning, and set the learning rate to 1 / 10 of the pre-training stage; Step S45: Inject Gaussian noise and random packet loss for adversarial enhancement training; Step S46: Remove the connections with weights less than 0.1 in the attention layer, and perform computational graph pruning; Step S47: Perform dynamic precision switching, use FP16 inference in the normal state, and automatically switch to FP32 precision when the prediction confidence is detected to decrease; Step S48: Use Kalman filtering to smooth the output delay sequence for the network prediction result.

6. The method for substation area situation awareness based on an intelligent analysis terminal according to claim 5, wherein In step S46, a dynamic gating mechanism is introduced in the multi-head attention layer, the connection channels with absolute weights <0.1 are automatically closed through learnable gating parameters, and Taylor expansion importance evaluation is used to calculate the influence degree of each attention head on the loss function; When initializing the global model, L1 sparse regularization is applied to the LSTM layer and the attention layer, and iterative progressive pruning is adopted. After every 10% pruning operation, 2 rounds of fine-tuning are performed until the model parameter quantity is reduced by 40% and the accuracy loss <3%.

7. A method for substation area situation awareness based on an intelligent analysis terminal according to claim 5, characterized in that In step S47, dynamic precision switching is performed, and model instances with two precisions, FP16 and FP32, are deployed, and weight parameters are shared through memory mapping technology; a precision switching channel is established, and a shared intermediate calculation result buffer is maintained in the GPU video memory to achieve zero-copy data transmission during precision mode switching; when seamless switching is performed, a hardware-accelerated precision conversion unit is used to automatically process the format alignment of FP16 to FP32; during the switching process, low-confidence samples are automatically reorganized into independent micro-batches, and when adaptive threshold adjustment is performed, the confidence threshold is dynamically adjusted based on the historical data distribution.

8. A method for substation area situation awareness based on an intelligent analysis terminal according to claim 1, characterized in that In step S6, the constructed performance evaluation indicators include the time alignment rate indicator and the maximum absolute value indicator; when constructing the digital twin test environment, a distributed sensor network is deployed to cover the key nodes of the physical system, and millisecond-level synchronous acquisition is achieved through the 5G / TSN protocol. Unstructured data sources are integrated, and the H.265 encoding is used to compress the transmission bandwidth to ensure that the spatio-temporal alignment error of multi-modal data is <50 ms; at the same time, based on the CAD model of the physical system, a three-dimensional digital model is reconstructed through point cloud scanning and deep learning to achieve sub-millimeter geometric accuracy; a multi-scale simulation engine is embedded to support the switching of test scenarios with different granularities.

9. A substation area situation awareness system based on an intelligent analysis terminal, comprising a security resilience data fusion unit and a security resilience awareness and warning unit, characterized in that: The security resilience data fusion unit includes a multi-source pre-purchase data acquisition module, a time scale alignment module, a feature extraction module, a security assessment and analysis module, and a fault prediction module; the multi-source pre-purchase data acquisition module is used to collect meteorological data, electrical data, and infrastructure monitoring data; the time scale alignment module is used to perform time scale alignment on structured data and unstructured data; The feature extraction module is used to extract features from the data with time scale alignment; the security assessment and analysis module is used to evaluate the trained model; the fault prediction module is used to predict faults in the substation area transformer; The security resilience awareness and warning unit includes network terminal devices, a big data analysis module, a real-time monitoring module, a deep learning module, an intelligent decision support system, an optimization scheduling module, a fault analysis module, and an intelligent repair module; the network terminal devices are used to collect substation area transformer data; the real-time monitoring module is used to perform real-time monitoring on the substation area transformer area; the deep learning module is used to predict communication delay through the LSTM network; the intelligent decision support system is used to provide decision-making assistance during the occurrence of substation area accidents; the optimization scheduling module is used to allocate personnel and resources after the accident occurs; the fault analysis module is used to perform fault analysis on the substation area transformer after the fault occurs; the intelligent repair module is used to generate a repair plan for the substation area transformer after the fault occurs.

10. A substation area situation awareness system based on an intelligent analysis terminal according to claim 9, characterized in that Before the accident occurs, the historical operation data and environmental parameters of the substation area transformer are mined by comprehensively using big data analysis technology and machine learning algorithms to predict potential fault hazards; under extreme event conditions, it replaces manual equipment inspection and real-time collects the status data of the equipment; During the accident process, real-time monitoring and emergency response are carried out. Deep learning and reinforcement learning are introduced to enhance the capabilities of the SCADA system, and real-time analysis and prediction of the status of distribution transformers are carried out; The intelligent decision support system provides the optimal emergency response plan; In the later stage of the accident, big data analysis technology and AI algorithms are used to carry out fault analysis and intelligent repair, identify fault patterns and root causes, automatically adjust the operating parameters of distribution transformers, and perform remote self-repair through the Internet of Things.

Images