A power transmission line snow cover monitoring device, method and system

Abstract

The application discloses a kind of power transmission line snow cover monitoring device, method and system, it is related to electric power system online monitoring technical field, the device is through the original weight of power transmission line snow cover and real-time environmental parameter, according to the data of acquisition executes power transmission line snow cover weighing zero drift adaptive compensation strategy, from physical layer, extract the multidimensional feature vector containing meteorological observation characteristics, physical wind pressure interference characteristics, time sequence derived characteristics, then utilize the LSTM time series regression model based on physical calculation and attention mechanism constructed, the zero drift of weighing sensor caused by multi-factor coupling is dynamically and accurately compensated, the real-time solution of snow cover weight is realized, and the accurate power transmission line snow cover weighing value is obtained.The application can realize the dynamic and accurate compensation of the zero drift of weighing sensor caused by multi-factor coupling, and improve the accuracy and reliability of snow cover weighing.

Description

Technical Field

[0001] This invention relates to the field of online monitoring technology for power systems, and in particular to a snow cover monitoring device, method and system for transmission lines. Background Technology

[0002] Snow accumulation on transmission lines can easily lead to safety accidents such as conductor galloping, tower tilting, and line breakage / tower collapse. Monitoring the weight of snow accumulation is a crucial aspect of power grid disaster prevention and mitigation. Weighing monitoring is widely used due to its intuitiveness and reliability, typically by installing load cells on transmission line towers to collect the weight of snow accumulation in real time. However, when load cells operate in the field for extended periods, they are susceptible to zero-point drift due to the combined effects of multiple factors such as temperature, humidity, and aging. This results in systematic deviations in snow accumulation weight monitoring, with errors being particularly significant under complex weather conditions such as strong winds and extreme low temperatures. This directly impacts the accuracy of snow accumulation warnings and power system operation and maintenance decisions.

[0003] Although weighing zero drift compensation strategies for snow cover monitoring of transmission lines have been proposed, existing solutions have many shortcomings: hardware compensation only targets the temperature characteristics of the sensor itself and cannot eliminate complex zero drift caused by wind pressure and dynamic environment; pure physical models are difficult to accurately characterize multi-factor coupled interference and cannot effectively distinguish between environmental interference and real load changes (such as dynamic pressure generated by wind pressure), which can easily cause compensation deviation; traditional machine learning methods have insufficient fitting ability, deep learning models have poor generalization performance, low compensation accuracy and easy failure under extreme weather conditions, and black box models have poor interpretability, which makes it difficult to meet the needs of actual monitoring. Summary of the Invention

[0004] To address the problem of measurement errors caused by zero-point drift of weighing sensors in existing snow-covered power transmission line weighing monitoring due to the coupling effect of multiple environmental factors, this invention provides a snow-covered power transmission line monitoring device, method, and system. The integrated monitoring device collects raw weight and meteorological environmental data, analyzes environmental interference from a physical perspective, and performs precise dynamic compensation for zero drift of the weighing sensor through an LSTM time-series regression model based on physical computation and attention mechanism.

[0005] In a first aspect, the present invention provides a snow-covered monitoring device for power transmission lines.

[0006] A snow cover monitoring device for power transmission lines, integrally installed on the transmission line tower, includes: The weighing acquisition unit includes a weighing pan, used to acquire the original weight of snow cover on the transmission line; The visual acquisition unit is used to acquire real-time images of the weighing pan; The micro-meteorological unit is deployed at the same location as the weighing and data acquisition unit to obtain real-time environmental parameters; The edge computing unit is used to preprocess the collected data, extract multi-dimensional feature vectors containing meteorological observation features, physical wind pressure interference features, and time-series derived features, output zero-drift prediction values ​​through a PA-LSTM-based prediction model, and realize real-time calculation of snow cover weight through dynamic compensation. The communication unit is used to upload the collected data and calculation results to the power system monitoring center cloud platform, and to receive the optimized model parameters issued by the cloud platform to update the prediction model in the edge computing unit.

[0007] Secondly, the present invention provides a method for monitoring snow cover on power transmission lines.

[0008] A method for monitoring snow cover on transmission lines, deployed in the edge computing unit of a snow cover monitoring device, enables zero-drift adaptive compensation for snow cover weighing on transmission lines, including: Real-time acquisition of raw snow weight and raw meteorological data at transmission line monitoring points; The raw meteorological data is analyzed and processed from a physical perspective to extract multi-dimensional feature vectors that include meteorological observation characteristics, physical wind pressure interference characteristics, and time-series derived characteristics. The multi-dimensional feature vector is input into the PA-LSTM-based prediction model, which passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs a zero-drift prediction value. Subtract the zero-drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero-drift compensation.

[0009] A further technical solution is that the raw meteorological data includes temperature, relative humidity, air pressure, wind speed, wind direction angle, and solar radiation intensity; The original weight was collected by a weighing pan installed at the monitoring point of the transmission line. The meteorological observation features are those reflecting the current basic meteorological conditions of the environment, including: instantaneous observations of temperature, relative humidity, air pressure, wind speed, wind direction angle, and solar radiation intensity after preprocessing; wherein, preprocessing includes outlier removal and standardization.

[0010] A further technical solution is that the physical wind pressure interference characteristics are generated based on a deterministic model of fluid dynamics laws, including the predicted wind pressure weight interference value, the component of the wind speed normal vector on the weighing pan surface, and the dynamic effective projected area that integrates the geometry of the weighing pan and the snow depth. Based on the principle of geometric projection, and combined with the side length of the weighing pan, the snow depth-side length equivalence coefficient, the snow depth of the weighing pan, and the empirical shape factor, the dynamic effective projected area that integrates the geometry of the weighing pan and the snow depth is calculated; the snow depth of the weighing pan is obtained by acquiring images of the weighing pan and using image recognition algorithms.

[0011] A further technical solution involves calculating the predicted wind pressure weight disturbance value as follows: Based on the real-time collected air pressure and temperature, the real-time air density is calculated by substituting it into the ideal gas law. Based on the real-time air density, combined with the preset force coupling coefficient, drag coefficient, and real-time collected wind speed and wind direction angle, the wind pressure calculation model is substituted to calculate the predicted wind pressure weight interference value.

[0012] A further technical solution is that the time-series derived features are calculated based on the original meteorological time-series data and are used to capture the dynamic change patterns of meteorology and zero drift, including: the rate of change of each meteorological parameter within a preset time interval and the mean and standard deviation within a set sliding time window.

[0013] A further technical solution, based on a PA-LSTM prediction model, takes multi-dimensional feature vectors as input and outputs zero-drift prediction values, including: Based on the feature embedding layer, the input meteorological observation features, physical wind pressure interference features, and time-series derived features are mapped to a unified semantic space to generate corresponding feature embedding vectors. Based on the feature-aware attention module, a learnable context query vector is introduced, using different feature embedding vectors as keys. Through the self-attention mechanism, the attention score of different features at the current time is calculated to obtain the feature representation after attention weighted fusion. Based on the temporal context encoding layer, the temporal features of the current time and historical time are represented, the temporal dependency of zero-point drift is captured, and the hidden state features of the current time are extracted. Based on the regression prediction layer, the final zero-point drift prediction value is output according to the hidden state features at the current time.

[0014] A further technical solution is that the training process of the PA-LSTM-based prediction model is as follows: The original meteorological data of the transmission line monitoring points under the baseline state and the true zero drift value of the snow-covered weight of the transmission line detected by the weighing sensor are obtained to form a time series dataset. The time series dataset is divided into training set and test set according to the time sequence. The baseline state is the state in which the transmission line has no additional snow load or the load is constant. The true zero drift value is the difference between the original snow-covered weight of the transmission line and the known constant reference weight. A PA-LSTM-based prediction model is trained using the training set, and the model parameters are optimized by minimizing the objective loss function during the training process to obtain the trained prediction model. The model is then tested using the test set.

[0015] In a further technical solution, the target loss function is a regularized total target loss function, which is a weighted sum of the mean square error term, the wind pressure disturbance amplitude constraint term, the L2 regularization term, and the attention sparsity L1 regularization term. The wind pressure disturbance amplitude constraint term, used to constrain the model's predicted values ​​within a reasonable range determined by physical principles, is expressed as: ; In the above formula, Indicates the number of samples. This is the zero drift value predicted by the model. This is the maximum wind pressure value. t This represents the sampling time, and the ReLU function is the activation function.

[0016] Further technical solutions include building an edge-cloud collaborative architecture, including: The edge computing unit, acting as an edge device, is used to locally cache the collected data and the compensated net snow weight of the transmission line, and then upload the data to the cloud. The power system monitoring center cloud platform is deployed in the cloud to store, integrate and manage the received data, automatically iterate and optimize the prediction model, and send the optimized model parameters to the edge terminal through an encrypted communication channel to complete the incremental update of the edge terminal model.

[0017] A further technical solution, wherein the automatic iterative optimization of the PA-LSTM model in the cloud includes: Acquire time-series data of snowless periods and historical snow-covered periods from multiple monitoring points uploaded from the edge terminal; the time-series data of snowless periods are continuous time period data with precipitation = 0, temperature > 0℃, and snow thickness = 0; Using snowless period time series data as the model calibration dataset, the PA-LSTM model is retrained using the calibration dataset according to a fixed period or a triggering strategy. During the training process, the optimal model parameters are solved by minimizing the total objective loss function. Using historical snow cover data as the test dataset, the accuracy of the trained PA-LSTM model is verified. If the evaluation index of the updated model is lower than that of the original model, the optimization is deemed effective. The optimized model parameters are compressed and packaged, and then sent to the edge in an incremental update manner through an encrypted channel. Otherwise, the optimization is deemed ineffective, and no update is performed.

[0018] Thirdly, the present invention provides a snow cover monitoring system for power transmission lines.

[0019] A snow cover monitoring system for transmission lines, integrated into a snow cover monitoring device for transmission lines, enables zero-drift adaptive compensation for snow cover weighing of transmission lines, comprising: The data acquisition module is used to acquire the original weight of snow cover and the original meteorological data at the monitoring points of the transmission lines in real time. The multi-dimensional feature processing module is used to analyze and process raw meteorological data from a physical perspective, extracting multi-dimensional feature vectors that include meteorological observation features, physical wind pressure interference features, and time-series derived features. The zero-drift prediction module is used to input multi-dimensional feature vectors into the PA-LSTM-based prediction model, which then passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs the zero-drift prediction value. The zero-drift adaptive compensation module is used to subtract the zero-drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero-drift compensation.

[0020] Fourthly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the above-described method for monitoring snow cover on power transmission lines when executing the executable instructions stored in the memory.

[0021] Fifthly, the present invention also provides a computer-readable storage medium storing executable instructions for causing a processor to execute the executable instructions to implement the above-described method for monitoring snow cover on power transmission lines.

[0022] In a sixth aspect, the present invention also provides a computer program product comprising executable instructions stored in a computer-readable storage medium; wherein, when the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the above-described method for monitoring snow cover on transmission lines.

[0023] The above one or more technical solutions have the following beneficial effects: This invention provides a snow-covered transmission line monitoring device, method, and system. By designing an integrated monitoring device, it collects raw weight and meteorological environmental data, analyzes environmental interference from a physical perspective, extracts multi-dimensional meteorological features, and utilizes a constructed Physics-InformedAttention Mechanism-Based LSTM Temporal Regression Model (PA-LSTM) to perform multi-dimensional feature fusion, dynamic feature attention weighting, temporal dependency capture, and accurate regression prediction. This enables dynamic and accurate prediction and compensation for zero-point drift of the weighing sensor caused by the coupling of multiple factors such as temperature, wind pressure, and humidity. This improves the accuracy, reliability, and interpretability of snow-covered transmission line weighing, and solves the problem of measurement errors caused by zero-point drift of the weighing sensor due to the coupling of multiple environmental factors in existing snow-covered transmission line weighing monitoring. In the PA-LSTM model, the physical model can finely model key interference factors such as wind pressure, thereby providing strong guidance that conforms to physical principles; the attention neural network captures the complex nonlinear residual drift generated by the coupling of multiple factors, which can achieve comprehensive high-precision compensation; through the above models, the interpretability, reliability and accuracy of snow cover weighing can be effectively improved. The present invention has high compensation accuracy, strong environmental adaptability, high interpretability and credibility and good engineering practicality, and is suitable for snow cover monitoring scenarios of transmission lines under complex meteorological conditions.

[0024] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 This is a schematic diagram of the structure of the snow cover monitoring device for transmission lines in Embodiment 1 of the present invention; Figure 2 This is an overall flowchart of the snow cover monitoring method for transmission lines in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the PA-LSTM-based prediction model in Embodiment 2 of the present invention.

[0027] Among them, 1. Weighing acquisition unit; 2. Visual acquisition unit; 3. Micro-meteorological unit. Detailed Implementation

[0028] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] As noted in the background section, snow weight measurement during transmission line snow cover monitoring suffers from zero drift, and existing weighing zero drift compensation schemes have many shortcomings: 1) Compensation methods based on the weighing sensor's own hardware: The built-in temperature compensation circuit can only compensate for the sensor's own temperature characteristics and cannot cope with complex zero-point drift caused by environmental factors. For example, in strong winds, this method is prone to misinterpreting the false increase in force caused by wind pressure as an increase in load, resulting in zero-drift compensation failure and making it difficult to adapt to complex meteorological environments; 2) Compensation methods based on physical models lack detailed modeling of environmental disturbances. For example, wind not only generates dynamic pressure but also causes secondary distribution of snow accumulation. Purely empirical models cannot effectively distinguish between environmental disturbances and actual load changes, easily leading to compensation bias. 3) Traditional compensation methods based on machine learning (such as linear regression) can only handle single or a small number of linear influencing factors. When faced with nonlinear dynamic drift caused by the coupling of multiple factors such as temperature, wind speed, and wind pressure, the model has insufficient fitting ability and poor compensation accuracy. On this basis, a prediction compensation method based on deep learning has been proposed. This prediction model is a pure data-driven "black box model" that only relies on data association to output prediction results. It lacks physical logic support, has poor interpretability, and has weak generalization ability. Under variable meteorological conditions, the influence weight of different factors on zero drift changes constantly, and it cannot adaptively adjust the influence weight, which leads to compensation failure under extreme weather conditions and makes it difficult to meet the needs of actual monitoring.

[0030] To address the aforementioned issues, this invention provides a snow-covered transmission line monitoring device, method, and system. By designing an integrated monitoring device, it collects raw weight and meteorological environmental data, analyzes environmental interference from a physical perspective, and constructs a physical computation and attention-based LSTM time-series regression model (PA-LSTM). This model meticulously models key interference factors such as wind pressure, providing strong guidance consistent with physical principles. Furthermore, an attention neural network captures complex nonlinear residual drift caused by multi-factor coupling, achieving comprehensive and high-precision compensation. Through multi-dimensional feature fusion, dynamic feature attention weighting, time-series dependency capture, and accurate regression prediction of this PA-LSTM model, dynamic and accurate prediction and compensation of zero-point drift of the weighing sensor caused by the coupling of multiple factors such as temperature, wind pressure, and humidity can be achieved. This improves the accuracy, reliability, and interpretability of zero-drift prediction, making it suitable for snow-covered transmission line monitoring scenarios under complex meteorological conditions.

[0031] Example 1 This embodiment proposes an integrated, harsh environment-resistant, and low-power transmission line snow cover monitoring device. The device adopts an outdoor industrial-grade tri-proof design and can work stably for a long time in complex field environments such as strong winds, blizzards, sandstorms, and salt spray. During installation, the device can be installed at a suitable location on the transmission line tower (usually no more than 5m away from the lowest conductor) according to the actual site conditions. The device can be used to synchronously, at the same point, and in real time monitor the snow cover weight, environmental weather, and snow depth of the transmission line.

[0032] Furthermore, the snow cover monitoring device proposed in this embodiment is composed of mechanical structure assembly, electrical function module, power supply and distribution module and other modules or components. The whole is modular integrated design with no exposed vulnerable parts. During installation, the device can be fixed on any crossbeam (or crossbar) of the tower body by using stainless steel clamps and high-strength insulating brackets. During installation, the weighing pan can be quickly adjusted to be horizontal by the three-dimensional tilt adjustment structure of the bracket, so that the weighing pan is strictly horizontal after installation, ensuring that the wind pressure and snow accumulation are evenly distributed.

[0033] like Figure 1 As shown, the snow cover monitoring device proposed in this embodiment mainly includes: (1) Weighing acquisition unit 1 includes a weighing pan, which serves as the core force and weight sensing component of the device. It is equipped with a pressure sensing component (or weighing sensor / sensing component) for high-precision acquisition of the original weight signal of snow cover on the transmission line. In this embodiment, the weighing acquisition unit adopts a square flat-side weighing pan to facilitate subsequent data processing operations.

[0034] As another optional implementation, the surface of the weighing pan is hard anodized and coated with a snow-repellent coating, making it non-stick and corrosion-resistant. Furthermore, the edge of the weighing pan can be provided with a 10mm high snow-blocking edge to prevent snow from sliding off, ensuring a stable effective measurement area, and the bottom of the snow-blocking edge is provided with several evenly distributed micro-drainage holes to facilitate automatic drainage after snow melts.

[0035] (2) Visual acquisition unit 2 is used to acquire images of the snow cover pattern on the weighing pan surface. It automatically identifies the snow depth using an edge AI algorithm (such as existing image recognition algorithms), providing accurate input for subsequent data processing, such as the dynamic effective projected area in wind pressure interference features. In this embodiment, the visual acquisition unit uses a camera, which can clearly image in low-light environments, and its field of view completely covers the weighing pan surface. As an optional implementation, the camera has a built-in infrared fill light that automatically turns on at night or in low visibility conditions during blizzards, without affecting the surrounding electromagnetic environment.

[0036] In this embodiment, the edge computing unit has a built-in lightweight snow depth recognition algorithm. This algorithm is used to complete defogging, noise reduction, and snow cover area segmentation locally, and the output snow depth time series data is aligned with the weight and meteorological data time.

[0037] (3) Micro-meteorological unit 3 is deployed at the same point as the weighing unit to ensure that meteorological data and weight data are synchronized in time and space, and is used to obtain all environmental parameters required for zero drift compensation. In this embodiment, the micro-meteorological unit includes sensors for temperature, humidity, pressure, wind speed, wind direction, solar radiation intensity, etc. All sensors adopt integrated meteorological probes, and the sampling frequency is strictly synchronized with the weight acquisition, with a default of 1 time / minute, which can be increased to 1 time / 10 seconds under extreme weather conditions.

[0038] (4) Edge computing unit, equipped with a zero-drift compensation model, is used to execute the zero-drift adaptive compensation strategy for snow cover weighing of transmission lines, and realize the real-time calculation of snow cover weight. Specifically, the edge computing unit is the core component of this device, equipped with a lightweight PA-LSTM zero-drift compensation model, and completes data acquisition, feature calculation, zero-drift compensation, local inference, and anomaly judgment, including: real-time acquisition of weighing, meteorological, and image data and preprocessing such as outlier removal and standardization; local calculation of wind pressure interference value, dynamic effective projected area, time-series derived features, etc., and extraction of multi-dimensional feature vectors containing meteorological observation features, physical wind pressure interference features, and time-series derived features; execution of zero-drift prediction and weight compensation, outputting the zero-drift prediction value through the PA-LSTM-based prediction model, and obtaining the net weight of snow cover after dynamic compensation, realizing the real-time calculation of snow cover weight; judging snow cover overload, sensor failure, communication anomaly, and triggering alarms locally.

[0039] (5) A communication unit, used to upload the collected data and calculation results to the power system monitoring center cloud platform, and to receive the optimized model parameters issued by the cloud platform to update the prediction model in the edge computing unit. In this embodiment, the transmission content of the communication unit includes: Regularly report: original weight, compensated net weight, meteorological data, snow depth, and equipment status; Incident reporting: Snow cover exceeding limits, sensor malfunction, low battery, communication interruption; Remote upgrade: Supports incremental updates of model parameters, with a single data transmission amount less than the set value, and does not consume bandwidth.

[0040] (6) Power supply unit, used to supply power to the entire device. This embodiment adopts a complementary power supply method of solar energy and lithium battery to ensure uninterrupted power supply during continuous rainy or snowy weather.

[0041] The specific workflow of the device proposed in this embodiment is as follows: After the entire device is powered on, each module, such as the weighing sensor, meteorological sensor, camera, battery, and communication module, performs a self-test; after the self-test is completed, raw weight, meteorological, and image data are synchronously collected at a set frequency; the edge computing unit extracts multi-dimensional physical features, inputs them into the PA-LSTM model to output zero-drift prediction values, and then calculates the predicted values ​​based on the raw weight. The zero-drift prediction value equals the net snow weight, and the final accurate net snow weight is calculated. The net snow weight, snow depth, weather conditions, and equipment status are encrypted and uploaded to the cloud platform. Simultaneously, when the net snow weight exceeds a threshold, an alarm message is immediately reported, and local audio-visual prompts are supported. Preferably, the device can receive optimized model parameters from the cloud and automatically update the local model.

[0042] Example 2 To address the problem of measurement errors caused by zero-point drift of weighing sensors in existing snow-covered transmission line monitoring systems due to the coupling effect of multiple environmental factors such as temperature, wind pressure, and humidity, this embodiment provides a snow-covered transmission line monitoring method. Deployed in the edge computing unit of the snow-covered transmission line monitoring device proposed in Embodiment 1, this method can accurately predict the zero-drift value of the weighing sensor (such as a weighing pan) under the current scenario. Then, the original weight of the snow cover (or snow accumulation) collected by the weighing sensor is compensated using the zero drift value, thereby realizing zero drift adaptive compensation for snow cover weighing of transmission lines, obtaining the true weight of the snow cover, and laying the foundation for accurate and reliable operation and maintenance decisions of the power system in the future.

[0043] The monitoring method proposed in this embodiment is as follows: Figure 2 As shown, the specific steps include: Step S1: Obtain the original weight of snow cover and original meteorological data at the transmission line monitoring point in real time.

[0044] Specifically, at monitoring points along transmission lines (such as below insulator strings on towers or at the midpoint of spans), the weighing pan of the device monitors the original weight of snow accumulation on the transmission line in real time and records the time. The original weight reading is Meanwhile, the system uses corresponding meteorological sensors to monitor raw meteorological data such as temperature, relative humidity, air pressure, wind speed, wind direction angle, and solar radiation intensity at the transmission line monitoring points in real time.

[0045] In one implementation, the weighing pan is square and horizontally positioned. It is used to collect the original weight of snow cover on the transmission lines at the monitoring point in real time at a preset sampling frequency (e.g., once per minute). Simultaneously, a micro-meteorological module collects raw meteorological data such as temperature, relative humidity, and wind speed at the monitoring point, with the sampling frequency consistent with the weighing data. A camera captures images of the weighing pan in real time, providing visual data for subsequent snow depth identification. The collected data is transmitted remotely via the device's built-in communication module, while also achieving initial local data caching.

[0046] By collecting the original weight of snow cover and multi-dimensional meteorological data in real time, the problem of single data source and spatiotemporal asynchrony in traditional schemes is solved. This ensures that the input data for zero drift prediction is comprehensive and time-series aligned. Moreover, multi-dimensional meteorological data can completely capture the core environmental factors affecting zero drift, avoiding prediction deviations caused by missing key data, and providing a data foundation for subsequent multi-factor coupled modeling.

[0047] Step S2: Analyze and process the raw meteorological data from a physical perspective, and extract multi-dimensional feature vectors that include meteorological observation characteristics, physical wind pressure interference characteristics, and time-series derived characteristics.

[0048] Specifically, considering that existing compensation schemes have problems such as hardware compensation being unable to cope with complex environmental interference, physical models lacking fine modeling, and traditional machine learning being unable to handle nonlinear drift, this embodiment adopts a deep learning strategy. However, a purely empirical model driven by data lacks physical logic support and has poor interpretability. Therefore, this embodiment starts from the physical level and conducts multi-faceted analysis on the raw meteorological data collected above, extracting multi-dimensional features containing physical characteristics to make the model features more reasonable and improve the interpretability of the model.

[0049] In this embodiment, the raw meteorological data is analyzed at the physical level to extract meteorological observation features, physical wind pressure interference features, and time-series derived features. The extracted features are then fused to form the input feature vector of the LSTM regression model. The extracted multi-dimensional feature vector It can be represented as: (1) In the above formula, Indicates the characteristics of original meteorological observations. This indicates the characteristics of physical wind pressure interference. Represents time-derived features, t This indicates the sampling time; the Concat function is a string concatenation function.

[0050] (1) Characteristics of primitive meteorological observations : In this embodiment, firstly, the real-time meteorological data collected from the transmission line monitoring points are formatted and standardized, including: removing outliers from the data, such as values ​​that exceed the reasonable physical range, to ensure data validity; and standardizing all data to avoid fitting calculation errors caused by different units.

[0051] Then, for the preprocessed multi-dimensional meteorological data, the preprocessed temperature... relative humidity air pressure Wind speed Wind angle and solar radiation intensity The instantaneous observation values ​​of these six meteorological parameters are combined in a fixed order to form a one-dimensional vector, which yields the meteorological observation characteristics, and can be expressed as: (2) The aforementioned meteorological observation characteristics can accurately reflect the basic meteorological conditions of the current environment, providing basic environmental characteristics to support subsequent zero-drift prediction.

[0052] It should be noted that all subsequent features are extracted based on the preprocessed meteorological data, and will not be elaborated further here.

[0053] (2) Characteristics of physical wind pressure interference : Based on a deterministic model of fluid dynamics laws, the interference of wind pressure on the weighing pan is quantified, generating physical wind pressure interference characteristics. These characteristics include the predicted wind pressure weight interference value, the component of the wind speed normal vector on the weighing pan surface, and the dynamic effective projected area integrating the weighing pan geometry and snow depth, which can be expressed as: (3) In the above formula, This represents the predicted wind pressure weight disturbance value. This represents the dynamic effective projected area that integrates the geometry of the weighing pan with the snow depth. This indicates the angle between the wind direction and the normal to the disk surface.

[0054] Specifically, regarding the predicted wind pressure weight interference value: based on the real-time collected air pressure and temperature, the real-time air density is calculated by substituting it into the ideal gas law. Then, based on this real-time air density, combined with the preset force coupling coefficient, drag coefficient, and the real-time collected wind speed and wind direction angle, it is substituted into the wind pressure calculation model to calculate the predicted wind pressure weight interference value. The calculation formula is as follows: (4) (5) In the above formula, Indicates the force coupling coefficient. Indicates real-time air density. Indicates the drag coefficient. Indicates air pressure. Indicates wind speed. This represents a specific gas constant for dry air. It represents the thermodynamic temperature of air.

[0055] Regarding the dynamic effective projected area of ​​the integrated weighing pan geometry and snow depth: Considering that the wind pressure experienced by the transmission line in the airflow direction is proportional to its projected area (i.e., effective projected area) on the plane perpendicular to the airflow direction, the dynamic effective projected area of ​​the integrated weighing pan geometry and snow depth is calculated based on the principle of geometric projection. The calculation formula is as follows: (6) In the above formula, Indicates the side length of the weighing pan. This represents the snow depth-side length equivalent coefficient. The snow depth of the weighing pan is indicated (this can be achieved by using a camera mounted on the detection equipment to capture an actual image of the weighing pan and identifying the snow depth through an image recognition algorithm). This represents the empirical shape factor.

[0056] Regarding the angle between the wind direction and the normal vector of the weighing pan: the angle between the wind direction and the normal vector of the weighing pan is obtained by using the tilt sensor on the equipment. The component of the wind speed on the normal vector of the pan is calculated by trigonometric functions, which can reflect the actual direction and intensity of the wind pressure on the sensor.

[0057] By employing the aforementioned feature extraction methods, wind pressure interference is quantified and refined at the physical level. This numerical and characterization of wind pressure interference provides strong guidance for subsequent model predictions, aligning with physical principles. Furthermore, the simultaneous integration of the dynamic effective projected area of ​​snow depth accurately reflects the actual degree of wind pressure interference to the sensor under snow cover conditions, improving the fit of wind pressure interference features. Unlike traditional purely empirical models, the features extracted from the physical level in this embodiment are more physically plausible, avoiding the poor interpretability of purely data-driven solutions.

[0058] (3) Temporal Derivative Features : Based on the preprocessed raw meteorological time-series data, the dynamic variation patterns of meteorological parameters and zero drift are calculated, and time-series derived features are extracted. These features include: the rate of change of each meteorological parameter within a preset time interval and the mean and standard deviation within a set sliding time window, which can be expressed as: (7) In the above formula, This indicates that various meteorological characteristics occur within a preset time interval. (like The rate of change within (minutes) can reflect the short-term dynamic changes of meteorological parameters; , In the sliding time window (like The characteristic mean and standard deviation calculated within minutes can reflect the medium- and long-term statistical regularities of meteorological parameters.

[0059] By capturing the temporal evolution patterns of meteorological parameters, such as trends like continuous temperature decreases and sudden wind speed increases, this approach addresses the static modeling shortcomings of traditional schemes that ignore the impact of historical environmental changes on current zero drift. Furthermore, by integrating short-term change rates with medium- and long-term statistical characteristics, it forms multi-scale temporal features. This allows subsequent models to perceive both real-time environmental conditions and capture long-term trends, providing rich dynamic information for feature encoding. This effectively addresses zero drift prediction in scenarios of sudden changes in meteorological parameters and improves the model's adaptability to extreme weather.

[0060] Step S3: Input the multi-dimensional feature vector into the PA-LSTM-based prediction model, which passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs the zero-drift prediction value.

[0061] The PA-LSTM model built in this embodiment is as follows: Figure 3 As shown, the model consists of four parts: feature embedding, feature-aware attention, temporal context encoding, and regression prediction. The multi-dimensional feature vector extracted in the previous step is input into this model, and the final output is a zero-drift prediction value. The process includes: Step S3.1: Based on the feature embedding layer, the input meteorological observation features, physical wind pressure interference features, and time-series derived features are mapped into a unified semantic space to generate corresponding feature embedding vectors.

[0062] Specifically, to avoid differences in feature spaces and dimensions among various feature classes, this embodiment uses a feature embedding layer to map meteorological, physical, and time-series features to a unified high-dimensional semantic space through learnable projection matrices, generating corresponding feature embedding vectors and achieving effective feature fusion. This mapping process can be represented as follows: (8) In the above formula, , For learnable projection matrix and bias, This represents a feature vector containing meteorological characteristics. This represents a feature vector that includes wind pressure disturbance characteristics. This represents a feature vector that contains time-derived features.

[0063] By using learnable projection matrices and biases, the model can automatically learn the spatial mapping rules of features, thereby improving the expressive power of features.

[0064] Step S3.2: Based on the feature-aware attention module, using different feature embedding vectors as keys, a learnable context query vector is introduced. Through the self-attention mechanism, the attention score of different features at the current time is calculated to obtain the feature representation after attention weighted fusion.

[0065] Specifically, considering the three features extracted above—meteorological, physical, and temporal—the impact of each feature on weighing drift varies over time in the actual environment. Traditional deep learning models are difficult to apply to this real-world scenario due to their weak generalization ability. To address this issue, this embodiment designs a feature-aware attention module, which dynamically calculates the current time... Attention scores for different features are assigned to perceive the weights of different features' influence on the model over time. The process is as follows: First, compute feature-aware attention. To simplify computation, each feature is embedded into a vector. It serves as a key-value pair and introduces a learnable contextual query vector. The self-attention mechanism is used to calculate the feature-aware attention of each feature. The calculation formula is as follows: (9) (10) In the above formula, This refers to the context query vector, which in this embodiment refers to... Query vector for the importance weight of a feature; express The first in Feature embedding vector, , That is, it corresponds to any one of the meteorological, physical, and temporal characteristics; This indicates that the calculation is based on the context query. and key value Attention score; , , These are the parameters for the attention network. Preferably, the above uses a query-based contextual attention mechanism, but a self-attention mechanism can also be used. This embodiment does not limit the specific attention mechanism.

[0066] Secondly, attention-weighted fusion is performed on the three types of embedded features to calculate the attention-weighted features. , can be represented as: (11) In the above formula, It is a feature of fusion Feature representation after attention weighting To reflect the dynamic perception of different types (i.e., meteorological, physical, and temporal characteristics), the influence weights on the model as time changes are determined.

[0067] Through the design of the aforementioned feature perception attention module, the attention scores of different features at the current moment can be dynamically calculated. It can automatically adjust the influence weights of different features according to real-time physical conditions (such as meteorological conditions), that is, adjust the influence weights of meteorological, physical, and temporal features in real time. For example, it can automatically increase the weight of wind pressure features under strong winds. It maintains robustness in complex environments where factors such as temperature, humidity, and wind speed change dynamically. It can effectively cope with the zero-point drift compensation requirements under extreme weather conditions and solve the problem that the fixed feature fusion logic of traditional solutions cannot adapt to changes in feature importance under extreme weather conditions. Its output attention weights can intuitively explain the basis of model decision-making, enabling operation and maintenance personnel to clearly understand the dominant influencing factors of current zero drift and further improve the credibility of the model in critical areas of power safety.

[0068] Step S3.3: Based on the temporal context encoding layer, represent the temporal features of the current time and historical time, capture the temporal dependency of zero-point drift, and extract the hidden state features of the current time.

[0069] Specifically, in the temporal context encoding layer, an existing LSTM network is used as the temporal encoder to continuously receive the output from the feature-aware attention module. Capture the temporal dependency of zero drift and extract the current time. The hidden state features can be represented as: (12) In the above formula, The feature vector after attention weighting. and They are time points Hidden states and unit values, These are the overall parameters of the LSTM.

[0070] In this embodiment, the hidden state is initialized. and unit value A zero vector represents an initial model with no historical memory. Subsequent time-series encoding is performed, whereby the vector is the result of the previous sampling time step (i.e., ...). The zero-point drift-related features extracted at time (time) are denoted as the hidden state. As the currently available short-term memory, the previous sampling time (i.e. The zero-point drift-related features accumulated at time (and all previous times) are denoted as the unit value. As a form of long-term memory, LSTM is based on short-term memory. and long-term memory , combined The feature vector at time step is updated to obtain Hidden state of time and unit value .

[0071] By leveraging the combined effects of long short-term and long-term memory, the model can effectively analyze previous weather changes and zero drift trends, avoiding the problems of gradient vanishing and inability to capture long-term temporal dependencies in traditional RNNs. This allows for accurate prediction of the current zero drift, making it suitable for scenarios where snow cover and zero drift change dynamically over time.

[0072] Step S3.4: Based on the regression prediction layer, output the final zero-point drift prediction value according to the hidden state features at the current moment.

[0073] Finally, the regression output from the regression prediction layer can be expressed as: (13) In the above formula, , These are output layer parameters, contained in ; This is the zero drift value predicted by the model.

[0074] Based on the above process, the prediction function of the PA-LSTM regression model can be described as follows: (14) In the above formula, It covers all learnable parameters of the embedding layer, attention network, LSTM, and output layer. Through a fully connected layer, high-dimensional temporal features are mapped to single-dimensional zero-drift prediction values, accurately outputting quantized zero-drift compensation criteria and avoiding prediction ambiguity caused by excessively high feature dimensions.

[0075] As a further technical solution, the training process of the prediction model based on PA-LSTM is as follows: First, the original meteorological data of the transmission line monitoring points under the baseline state and the real zero drift value of the snow cover weight of the transmission line detected by the weighing sensor are obtained to form a time series dataset. The time series dataset is then divided into a training set and a test set according to the time sequence.

[0076] Specifically, model training and evaluation are based on a time-aligned sequence dataset collected under baseline conditions (no additional snow cover load or constant load), which can be represented as: ; In the above formula, Represents the set of sample points indexed in chronological order; Indicates at time place 3D input feature vector; At any moment The true zero drift value at the specified location is used as a label for supervised learning. This true zero drift value is the difference between the original weight of the snow cover on the transmission line and the known constant reference weight. ,in For a moment The original weight reading, The reference weight is a known constant.

[0077] Preferably, the original weight readings of the weighing sensor and the known constant reference weight are extracted during the snowless period. The true zero drift value is calculated according to the formula: True zero drift = Original weight - Reference weight. These true zero drift values ​​are then used to train the regression prediction layer parameters of the PA-LSTM model, correcting the model's prediction bias for the sensor's own zero drift and solving the problem of zero drift reference changes caused by sensor aging and wire creep. At the same time, the snowless period data contains different meteorological conditions (such as low temperature and no wind, high temperature and high humidity, strong wind, etc.). Using this data to expand the training set can enable the model to learn the zero drift rules under more extreme weather scenarios, improving the model's generalization ability. Moreover, for the strong wind weather data during the snowless period, the calculation parameters of physical wind pressure interference characteristics (such as force coupling coefficient and drag coefficient) can be optimized to improve the model's compensation accuracy under strong wind interference during the snow-covered period.

[0078] Furthermore, the data in the time series dataset is partitioned (i.e., time series segmentation) as follows: ; Among them, symbols This means that the test set is strictly divided according to chronological order to ensure that it represents future, unseen weather patterns, prevent information leakage, and evaluate the model's true generalization ability. , , , These are the training set, validation set, test set, and total dataset, respectively.

[0079] Next, a PA-LSTM-based prediction model is trained using the training set. During the training process, the model parameters are optimized by minimizing the objective loss function to obtain the trained prediction model, which is then tested using the test set.

[0080] In this embodiment, the objective loss function is designed as a weighted sum of the mean square error term, the wind pressure disturbance amplitude constraint term, the L2 regularization term, and the attention sparsity L1 regularization term. The final regularized overall objective loss function can be expressed as: (15) In the above formula, , , The weights of each loss term, This is the mean square error term. For L2 regularization terms, For attention sparsity L1 regularization, This is a constraint term for the amplitude of wind pressure interference.

[0081] Furthermore, the aforementioned mean squared error term is used to minimize the numerical deviation between the model's predicted zero drift value and the actual zero drift value. It is a fundamental core term of the loss function, directly determining the model's basic prediction accuracy for the zero drift value and ensuring that the zero drift prediction value output by the model is as close as possible to the actual value. This term can be expressed as: (16) In addition, the L2 regularization term is used to apply a squared penalty to all learnable parameters of the model to prevent overfitting and improve the model's generalization ability. Attention sparsity L1 regularization term is used to penalize the feature attention weights sparsely, encouraging the model's attention weights to tend towards a sparse state (i.e. most weights are close to 0, and a few key feature weights are significant), so that the model learns to "focus" on key features, improves the model's feature learning efficiency, and outputs clear feature attention weights, making it easier to intuitively understand the model's decision-making basis and improve the model's interpretability. The wind pressure disturbance amplitude constraint term is used to constrain the zero-drift prediction value of the model within the physically reasonable range determined by fluid dynamics principles. This avoids unreasonable prediction results that exceed physical laws due to pure data fitting, ensuring that the data-driven model output conforms to physical logic and guaranteeing the rationality of the prediction results from the bottom layer. This term can be expressed as: (17) In the above formula, Indicates the number of samples. This is the zero drift value predicted by the model. This represents the maximum wind pressure value.

[0082] By designing the above objective loss function, we can ensure that the deviation between the predicted value and the actual zero drift is minimized, prevent the model from overfitting, encourage the model to focus on key features, improve generalization ability, and solve the problem of model compensation failure under extreme weather conditions.

[0083] During the training process described above, parameter optimization can be expressed as: (18) The optimal parameter set can be solved using optimization algorithms such as the Adam algorithm. This yields the trained prediction model.

[0084] Step S4: Subtract the zero drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero drift compensation.

[0085] Specifically, a direct subtraction compensation logic is adopted to calculate the time step. t The net weight of snow cover on the transmission line after zero-drift compensation is obtained by subtracting the zero-drift predicted value from the PA-LSTM model's output from the original weight of snow cover on the transmission line. The calculation formula is: (19) In the above formula, For a moment The original weight reading, for The zero drift value predicted by PA-LSTM at time step.

[0086] As a further implementation method, based on the accurate net weight value of snow cover after zero drift compensation, the snow cover thickness can be calculated by combining known parameters such as the weight of the transmission line itself and snow density. Furthermore, the net weight value and / or thickness of the snow cover can be used to further assist in judging the risk of tilting and falling of transmission line towers. For example, when the net weight value of snow cover exceeds a preset threshold, a tower tilt risk alarm is triggered. By providing early warning of tower safety hazards caused by snow cover, the safe and stable operation of transmission lines can be ensured, and data support can be provided for relevant operation and maintenance decisions.

[0087] By introducing a feature-aware attention mechanism into the PA-LSTM model through the method described in this embodiment, the model possesses dynamic feature selection capabilities. It can automatically adjust the influence weights of different features according to real-time meteorological conditions, maintaining robustness in complex environments with dynamic changes in factors such as temperature, humidity, and wind speed. It effectively addresses the zero-point drift compensation requirements under extreme weather conditions, enabling dynamic and accurate prediction and compensation for zero-point drift of weighing sensors caused by the coupling of multiple factors such as temperature, wind pressure, and humidity, thereby improving the accuracy of snow-covered weighing of transmission lines. While outputting the zero-point drift prediction value, the model can also output the attention weights and uncertainty estimates of each feature, allowing maintenance personnel to intuitively understand the model's decision-making basis. For example, the model's attention to wind speed features under strong winds can enhance the credibility of the system's application in critical areas of power safety, providing data support for power grid-related maintenance decisions.

[0088] As another implementation method, this embodiment establishes an edge-cloud collaborative architecture, in which the edge computing unit, as the edge end, is implemented through a low-power processor and directly connects to the weighing sensor, meteorological sensor and tilt sensor; the cloud platform of the provincial / regional monitoring center of the power system is deployed in the cloud, relying on high-performance server clusters, distributed storage systems, etc. to provide sufficient computing and storage resources.

[0089] Based on the above design, the edge terminal is used to execute the above-mentioned zero-drift adaptive compensation strategy for snow-covered weighing of transmission lines, and realize real-time compensation response. That is, the edge computing node connects to the meteorological sensor and deploys a lightweight PA-LSTM model to collect and preprocess the sensor data in real time. According to the preprocessed multi-dimensional feature vector, it quickly performs feature embedding, attention weighting, time-series coding and regression prediction, and calculates the compensated snow-covered net weight according to formula (19), i.e., compensated net weight = original weight - zero-drift prediction value, thereby completing the real-time inference calculation. In addition, the edge terminal will also upload the collected data and calculation results to the cloud monitoring center in real time, and at the same time cache the compensation data of the most recent 24 hours locally to prevent transmission interruption and loss. Preferably, if the edge terminal detects that the model prediction error exceeds the threshold (e.g., ≥5%) or the sensor data is abnormal, it will immediately send an alarm message to the cloud and activate the backup compensation strategy (e.g., using the historical average zero-drift value for compensation) to ensure business continuity.

[0090] The cloud-based system stores and integrates data uploaded from the edge nodes, supports efficient querying, and performs automatic iterative optimization of the PA-LSTM model. The optimized model parameters are then distributed to the edge nodes via an encrypted channel. Specifically, the cloud automatically identifies snowless periods, such as consecutive time periods in meteorological data where "precipitation = 0, temperature > 0℃, snow depth = 0," and selects the time-series data from these periods (i.e., snowless period time-series data) as the model calibration dataset. No manual annotation is required; model training is periodically triggered (or triggered under other conditions) using this data, enabling automatic iterative optimization. Preferably, a fixed training cycle (every 3 months) or triggered training (e.g., setting a trigger condition: when the average prediction error reported by multiple edge nodes > 3%) can be set, and the cloud automatically starts the training task.

[0091] Specifically, the process of automatic iterative optimization of the cloud model includes: acquiring snowless period time-series data (i.e., continuous time-series data with precipitation = 0, temperature > 0℃, and snow thickness = 0) and historical snow cover period data from multiple monitoring points uploaded from the edge; using the snowless period time-series data as the calibration dataset for the model; dividing the calibration dataset into training and testing sets; starting model training according to a fixed period or triggering strategy; retraining the PA-LSTM model using the training set; and solving for the optimal parameters of the model using the Adam optimization algorithm. The entire training process is automated and requires no manual intervention.

[0092] After training, the retrained PA-LSTM model is evaluated on the test set by evaluating metrics such as the coefficient of determination. To evaluate model performance, if the updated model's evaluation metric value is lower than the original model, the optimization is considered effective; otherwise, the training result is discarded. Alternatively, to avoid frequent model updates due to minor performance optimizations, a threshold can be pre-set. If the decrease in the evaluation metric value exceeds the set threshold, the optimization is considered effective; otherwise, the training result is discarded.

[0093] Finally, when the optimization is deemed effective, the updated optimal model parameters (projection matrix, attention network parameters, LSTM weights, etc.) are compressed and packaged, and sent to the corresponding edge nodes through an encrypted channel. The old parameters of the edge model can be replaced by incremental updates, and the amount of data transmitted is controlled at the KB level to avoid consuming too much bandwidth.

[0094] Through the aforementioned edge-cloud collaborative architecture, the edge side ensures real-time compensation response, while the cloud side enables automatic iterative optimization of the model and can utilize snowless period data to complete model calibration. This eliminates the need for extensive manual intervention, reduces maintenance costs, and facilitates engineering deployment and promotion.

[0095] Example 3 This embodiment provides an adaptive compensation system for zero drift in snow-covered weighing of transmission lines, which is integrated into the snow-covered monitoring device for transmission lines proposed in Embodiment 1. This system enables adaptive compensation for zero drift in snow-covered weighing of transmission lines. The system includes: The data acquisition module is used to acquire the original weight of snow cover and the original meteorological data at the monitoring points of the transmission lines in real time. The multi-dimensional feature processing module is used to analyze and process raw meteorological data from a physical perspective, and extract multi-dimensional feature vectors that include meteorological observation features, physical wind pressure interference features, and time-series derived features. The zero-drift prediction module is used to input multi-dimensional feature vectors into the PA-LSTM-based prediction model, which then passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs the zero-drift prediction value. The zero-drift adaptive compensation module is used to subtract the zero-drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero-drift compensation.

[0096] Example 4 This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0097] Example 5 This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0098] Example 6 This embodiment provides a computer program product including executable instructions, which are computer instructions; the executable instructions are stored in a computer-readable storage medium. When the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the electronic device performs the method described in this embodiment.

[0099] The steps and methods involved in Embodiments 3 to 6 above correspond to those in Embodiment 2. For specific implementation details, please refer to the relevant description section of Embodiment 2. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0100] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0101] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. A snow cover monitoring device for power transmission lines, characterized in that, The entire assembly is installed on the transmission line tower, including: The weighing acquisition unit includes a weighing pan, used to acquire the original weight of snow cover on the transmission line; The visual acquisition unit is used to acquire real-time images of the weighing pan; The micro-meteorological unit is deployed at the same location as the weighing and data acquisition unit to obtain real-time environmental parameters; The edge computing unit is used to preprocess the collected data, extract multi-dimensional feature vectors containing meteorological observation features, physical wind pressure interference features, and time-series derived features, output zero-drift prediction values ​​through a PA-LSTM-based prediction model, and realize real-time calculation of snow cover weight through dynamic compensation. The communication unit is used to upload the collected data and calculation results to the power system monitoring center cloud platform, and to receive the optimized model parameters issued by the cloud platform to update the prediction model in the edge computing unit. The multi-dimensional feature vector is input into the PA-LSTM-based prediction model, which outputs zero-drift prediction values, including: Based on the feature embedding layer, the input meteorological observation features, physical wind pressure interference features, and time-series derived features are mapped to a unified high-dimensional semantic space through a learnable projection matrix, generating corresponding feature embedding vectors. Based on the feature-aware attention module, different feature embedding vectors are used. For key-value pairs, a learnable contextual query vector is introduced. Through a self-attention mechanism, the attention scores and feature-aware attention of different features at the current time are calculated. Attention-weighted fusion is then performed on the three types of embedded features to obtain the attention-weighted fused feature representation, as follows: ； ； ； In the above formula, Represents the context query vector; Indicates the first Feature embedding vector, , This corresponds to any one of the following characteristics: meteorological, physical, and temporal. This indicates that the calculation is based on the context query. and key value Attention score; Represents feature perception attention; , , These are the parameters of the attention network; It is a feature of fusion Feature representation after attention weighting; Based on the temporal context encoding layer, the temporal features of the current time and historical time are represented, the temporal dependency of zero-point drift is captured, and the hidden state features of the current time are extracted. Based on the regression prediction layer, the final zero-point drift prediction value is output according to the hidden state features at the current time.

2. A method for monitoring snow cover on power transmission lines, characterized in that, The edge computing unit deployed in the snow cover monitoring device for transmission lines as described in claim 1 realizes zero-drift adaptive compensation for snow cover weighing of transmission lines, including: Real-time acquisition of raw snow weight and raw meteorological data at transmission line monitoring points; The raw meteorological data is analyzed and processed from a physical perspective to extract multi-dimensional feature vectors that include meteorological observation characteristics, physical wind pressure interference characteristics, and time-series derived characteristics. The multi-dimensional feature vector is input into the PA-LSTM-based prediction model, which passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs a zero-drift prediction value. Subtract the zero drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero drift compensation. The multi-dimensional feature vector is input into the PA-LSTM-based prediction model, which outputs zero-drift prediction values, including: Based on the feature embedding layer, the input meteorological observation features, physical wind pressure interference features, and time-series derived features are mapped to a unified high-dimensional semantic space through a learnable projection matrix, generating corresponding feature embedding vectors. Based on the feature-aware attention module, different feature embedding vectors are used. For key-value pairs, a learnable contextual query vector is introduced. Through a self-attention mechanism, the attention scores and feature-aware attention of different features at the current time are calculated. Attention-weighted fusion is then performed on the three types of embedded features to obtain the attention-weighted fused feature representation, as follows: ； ； ； In the above formula, Represents the context query vector; Indicates the first Feature embedding vector, , This corresponds to any one of the following characteristics: meteorological, physical, and temporal. This indicates that the calculation is based on the context query. and key value Attention score; Represents feature perception attention; , , These are the parameters of the attention network; It is a feature of fusion Feature representation after attention weighting; Based on the temporal context encoding layer, the temporal features of the current time and historical time are represented, the temporal dependency of zero-point drift is captured, and the hidden state features of the current time are extracted. Based on the regression prediction layer, the final zero-point drift prediction value is output according to the hidden state features at the current time.

3. The method for monitoring snow cover on transmission lines as described in claim 2, characterized in that, The raw meteorological data includes temperature, relative humidity, air pressure, wind speed, wind direction angle, and solar radiation intensity; The original weight was collected by a weighing pan installed at the monitoring point of the transmission line. The meteorological observation features are those reflecting the current basic meteorological conditions of the environment, including: instantaneous observations of temperature, relative humidity, air pressure, wind speed, wind direction angle, and solar radiation intensity after preprocessing; wherein, preprocessing includes outlier removal and standardization.

4. The method for monitoring snow cover on transmission lines as described in claim 2, characterized in that, The physical wind pressure interference characteristics are generated based on a deterministic model of fluid dynamics laws, including the predicted wind pressure weight interference value, the component of the wind speed normal vector on the weighing pan surface, and the dynamic effective projected area that integrates the weighing pan geometry and snow depth. Based on the principle of geometric projection, and combined with the side length of the weighing pan, the snow depth-side length equivalence coefficient, the snow depth of the weighing pan, and the empirical shape factor, the dynamic effective projected area that integrates the geometry of the weighing pan and the snow depth is calculated; the snow depth of the weighing pan is obtained by acquiring images of the weighing pan and using image recognition algorithms.

5. The method for monitoring snow cover on transmission lines as described in claim 4, characterized in that, The calculation process for predicting wind pressure weight disturbance values ​​is as follows: Based on the real-time collected air pressure and temperature, the real-time air density is calculated by substituting it into the ideal gas law. Based on the real-time air density, combined with the preset force coupling coefficient, drag coefficient, and real-time collected wind speed and wind direction angle, the wind pressure calculation model is substituted to calculate the predicted wind pressure weight interference value.

6. The method for monitoring snow cover on transmission lines as described in claim 2, characterized in that, The time-series derived features are calculated based on the original meteorological time-series data and are used to capture the dynamic changes in meteorology and zero drift, including: the rate of change of each meteorological parameter within a preset time interval and the mean and standard deviation within a set sliding time window.

7. The method for monitoring snow cover on transmission lines as described in claim 2, characterized in that, The training process of the PA-LSTM-based prediction model is as follows: The original meteorological data of the transmission line monitoring points under the baseline state and the true zero drift value of the snow-covered weight of the transmission line detected by the weighing sensor are obtained to form a time series dataset. The time series dataset is divided into training set and test set according to the time sequence. The baseline state is the state in which the transmission line has no additional snow load or the load is constant. The true zero drift value is the difference between the original snow-covered weight of the transmission line and the known constant reference weight. A PA-LSTM-based prediction model is trained using the training set, and the model parameters are optimized by minimizing the objective loss function during the training process to obtain the trained prediction model. The model is then tested using the test set.

8. The method for monitoring snow cover on transmission lines as described in claim 7, characterized in that, The target loss function is a regularized total target loss function, which is a weighted sum of the mean square error term, the wind pressure disturbance amplitude constraint term, the L2 regularization term, and the attention sparsity L1 regularization term.

9. The method for monitoring snow cover on transmission lines as described in claim 8, characterized in that, The wind pressure disturbance amplitude constraint term, used to constrain the model's predicted values ​​within a reasonable range determined by physical principles, is expressed as: ； In the above formula, Indicates the number of samples. This is the zero drift value predicted by the model. This is the maximum wind pressure value. t This represents the sampling time, and the ReLU function is the activation function.

10. The method for monitoring snow cover on transmission lines as described in claim 2, characterized in that, Building an edge-cloud collaborative architecture includes: The edge computing unit, acting as an edge device, is used to locally cache the collected data and the compensated net snow weight of the transmission line, and then upload the data to the cloud. The power system monitoring center cloud platform is deployed in the cloud to store, integrate and manage the received data, automatically iterate and optimize the prediction model, and send the optimized model parameters to the edge terminal through an encrypted communication channel to complete the incremental update of the edge terminal model.

11. The method for monitoring snow cover on transmission lines as described in claim 10, characterized in that, The cloud-based automatic iterative optimization of the PA-LSTM model includes: Acquire time-series data of snowless periods and historical snow-covered periods from multiple monitoring points uploaded from the edge terminal; the time-series data of snowless periods are continuous time period data with precipitation = 0, temperature > 0℃, and snow thickness = 0; Using snowless period time series data as the model calibration dataset, the PA-LSTM model is retrained using the calibration dataset according to a fixed period or a triggering strategy. During the training process, the optimal model parameters are solved by minimizing the total objective loss function. Using historical snow cover data as the test dataset, the accuracy of the trained PA-LSTM model is verified. If the evaluation index of the updated model is lower than that of the original model, the optimization is deemed effective. The optimized model parameters are compressed and packaged, and then sent to the edge in an incremental update manner through an encrypted channel. Otherwise, the optimization is deemed ineffective, and no update is performed.

12. A snow cover monitoring system for power transmission lines, characterized in that, Equipped in the snow cover monitoring device for transmission lines as described in claim 1, it achieves zero-drift adaptive compensation for snow cover weighing of transmission lines, including: The data acquisition module is used to acquire the original weight of snow cover and the original meteorological data at the monitoring points of the transmission lines in real time. The multi-dimensional feature processing module is used to analyze and process raw meteorological data based on the physical level, and extract multi-dimensional feature vectors containing meteorological observation features, physical wind pressure interference features, and time-series derived features. The zero-drift prediction module is used to input multi-dimensional feature vectors into the PA-LSTM-based prediction model, which then passes through the feature embedding layer, the feature-aware attention module, the temporal context encoding layer, and the regression prediction layer in sequence, and outputs the zero-drift prediction value. The zero-drift adaptive compensation module is used to subtract the zero-drift prediction value from the original weight of the snow covering the transmission line to obtain the net weight of the snow covering the transmission line after zero-drift compensation. The multi-dimensional feature vector is input into the PA-LSTM-based prediction model, which outputs zero-drift prediction values, including: Based on the feature embedding layer, the input meteorological observation features, physical wind pressure interference features, and time-series derived features are mapped to a unified high-dimensional semantic space through a learnable projection matrix, generating corresponding feature embedding vectors. Based on the feature-aware attention module, different feature embedding vectors are used. For key-value pairs, a learnable contextual query vector is introduced. Through a self-attention mechanism, the attention scores and feature-aware attention of different features at the current time are calculated. Attention-weighted fusion is then performed on the three types of embedded features to obtain the attention-weighted fused feature representation, as follows: ； ； ； In the above formula, Represents the context query vector; Indicates the first Feature embedding vector, , This corresponds to any one of the following characteristics: meteorological, physical, and temporal. This indicates that the calculation is based on the context query. and key value Attention score; Represents feature perception attention; , , These are the parameters of the attention network; It is a feature of fusion Feature representation after attention weighting; Based on the temporal context encoding layer, the temporal features of the current time and historical time are represented, the temporal dependency of zero-point drift is captured, and the hidden state features of the current time are extracted. Based on the regression prediction layer, the final zero-point drift prediction value is output according to the hidden state features at the current time.

13. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the transmission line snow cover monitoring method according to any one of claims 2-11.

14. A computer-readable storage medium, characterized in that, The device stores executable instructions that, when executed by a processor, implement the snow cover monitoring method for transmission lines as described in any one of claims 2-11.

15. A computer program product, characterized in that, The computer program product includes executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the transmission line snow cover monitoring method according to any one of claims 2-11.

Images