A power transmission conductor-based temperature early warning method and system

Abstract

The present application relates to the technical field of power transmission conductor temperature monitoring and early warning, and discloses a temperature early warning method and system based on a power transmission conductor. The method collects multi-source temperature monitoring information of the power transmission conductor. The multi-source temperature monitoring information is processed to extract temperature change characteristics, including spatial temperature gradient distribution and time series trend, in order to more deeply analyze the conductor temperature change. The extracted temperature change characteristics are analyzed by a pre-trained temperature anomaly detection model to generate a set of potential hot spot areas, which are associated with temperature exceeding probability and geographic coordinates, and accurately locate the potential risk areas. The set of potential hot spot areas and real-time environmental parameters are fused to perform multi-modal data collaborative evaluation, generate a comprehensive risk feature vector, and realize comprehensive risk assessment. The comprehensive risk feature vector is processed by a warning decision engine to output a temperature early warning instruction containing a warning level and a dynamic adjustment parameter.

Description

Technical Field

[0001] This invention relates to the field of temperature monitoring and early warning technology for power transmission lines, specifically to a temperature early warning method and system based on power transmission lines. Background Technology

[0002] With the continuous advancement of industrial technology and the vigorous development of the social economy, the demand for electricity from all sectors is growing rapidly. This trend has led to the continuous expansion of transmission lines and the increasing load they bear. As a key component of power transmission, the stability of the operating state of transmission lines directly affects the safe and reliable operation of the entire power grid. Under high-voltage and high-current operating conditions, transmission lines are prone to generating heat due to the thermal effect of current. If their temperature is not monitored and controlled in a timely and effective manner, excessively high temperatures can cause a decline in the mechanical properties of the conductors, an increase in sag, and even serious faults such as short circuits and line breaks, leading to large-scale power outages and causing significant negative impacts on social production and daily life. For example, during the high-temperature period in summer, the surge in electricity load often causes a rapid rise in the temperature of transmission lines. If monitoring is inadequate, the aforementioned faults can easily occur, affecting residential electricity use and industrial production. Therefore, accurate monitoring of transmission line temperature has become a crucial link in ensuring the safe and stable operation of the power grid.

[0003] Traditional power transmission conductor temperature monitoring often focuses solely on monitoring the conductor's surface temperature. However, this method has several limitations. Firstly, complex and variable external environmental factors, such as direct sunlight and wind and rain, can interfere with surface temperature measurements, significantly reducing the accuracy of the results and failing to accurately reflect the actual internal heating of the conductor. Secondly, relying solely on surface temperature cannot comprehensively assess the conductor's actual operating condition under different load conditions, and it cannot promptly identify potential safety hazards.

[0004] Besides the limitations of traditional monitoring methods, traditional manual inspection also has significant shortcomings. Manual inspection is not only inefficient but also heavily influenced by the professional level and subjective factors of the inspectors. Faced with an increasingly vast transmission line network, manual inspection struggles to be comprehensive and detailed, failing to detect subtle changes in conductor temperature and potential safety hazards. In practice, manual inspection often consumes substantial manpower, resources, and time, and cannot achieve real-time monitoring. If an abnormal rise in conductor temperature occurs during inspection intervals, it is difficult to detect and address in time, potentially leading to serious line faults. According to incomplete statistics, transmission line faults caused by abnormal conductor temperature result in enormous economic losses for power companies annually and severely disrupt normal social production and daily life.

[0005] To address the shortcomings of existing transmission line temperature monitoring methods and meet the urgent needs of modern intelligent transmission line management, it is essential to propose a novel, more scientific, and effective temperature early warning method. This new method needs to comprehensively consider multiple factors, overcome the limitations of traditional monitoring methods, and achieve comprehensive, high-precision monitoring and early warning of transmission line temperature. By introducing multi-source temperature monitoring information, it comprehensively captures the temperature change characteristics of the conductors. Utilizing advanced detection models and data analysis techniques, it can promptly and accurately identify potential hotspots and conduct a comprehensive assessment based on real-time environmental parameters. This provides maintenance personnel with more accurate and reliable decision-making support, effectively preventing transmission line faults caused by abnormal temperatures, ensuring the safe and stable operation of the power grid, and meeting society's ever-increasing demands for reliable power supply. Summary of the Invention

[0006] The purpose of this invention is to provide a temperature early warning method and system based on power transmission lines to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a temperature early warning method based on power transmission lines, the method comprising:

[0008] Collect multi-source temperature monitoring information of power transmission lines, including infrared thermal image data, contact sensor readings, and environmental parameters;

[0009] The multi-source temperature monitoring information is processed to extract temperature change features, which include spatial temperature gradient distribution and time series trends.

[0010] The temperature change characteristics are analyzed by a pre-trained temperature anomaly detection model to generate a set of potential hotspot areas, which are associated with the probability of temperature exceeding the standard and geographical coordinates.

[0011] By integrating the set of potential hotspot areas with real-time environmental parameters, a multimodal data collaborative assessment is performed to generate a comprehensive risk feature vector;

[0012] The comprehensive risk feature vector is processed in a hierarchical manner using an early warning decision engine, and a temperature early warning command is output. The temperature early warning command includes the early warning level and dynamic adjustment parameters.

[0013] Preferably, the multi-source temperature monitoring information collected from the transmission line includes:

[0014] Continuous infrared scanning of the transmission line was performed to obtain infrared thermal images and their corresponding timestamp sequences;

[0015] Synchronously read data from contact temperature sensors arranged on the surface of the conductors to generate a sensor reading stream;

[0016] Collect environmental parameter information, including ambient temperature, humidity, wind speed, and solar radiation intensity;

[0017] The infrared thermal image, sensor reading stream, and environmental parameter information are time-aligned to generate a time-synchronized multi-source data set.

[0018] The multi-source data set is subjected to noise filtering and missing value imputation to generate a standardized monitoring dataset.

[0019] Preferably, the process of processing the multi-source temperature monitoring information to extract temperature change features includes:

[0020] The standardized monitoring dataset is divided into spatial grids to generate multiple temperature analysis regions;

[0021] Calculate the magnitude and direction of the temperature gradient within each temperature analysis region to generate a spatial gradient distribution map;

[0022] Extract time series trend features, including temperature change rate, periodic fluctuations, and abnormal fluctuation patterns;

[0023] The spatial gradient distribution map and the time series trend features are tensor-concatenated to generate a multidimensional temperature change feature matrix.

[0024] Preferably, the analysis of the temperature change characteristics using a pre-trained temperature anomaly detection model includes:

[0025] The multidimensional temperature change feature matrix is ​​input into a convolutional neural network for local feature extraction to generate a feature activation map.

[0026] The feature activation map is processed by a region proposal network to generate candidate anomaly regions and their confidence scores;

[0027] The nonmaximum suppression algorithm is applied to filter the candidate abnormal regions and remove regions with high overlap.

[0028] A set of potential hotspot regions is generated based on a confidence threshold. Each potential hotspot region includes its center coordinates and the probability of temperature anomalies.

[0029] Preferably, the fusion of the potential hotspot area set and real-time environmental parameters includes:

[0030] Based on the center coordinates of the set of potential hotspot areas, extract the corresponding subset of environmental parameters;

[0031] The environmental parameter subset is feature-encoded to generate an environmental feature vector;

[0032] The environmental feature vector is weighted and fused with the temperature anomaly probability of the potential hotspot area set;

[0033] The contribution weights of environmental factors to temperature anomalies are calculated using an attention mechanism to generate a weighted fusion result.

[0034] The weighted fusion result is correlated and mapped with the geographical coordinates of the potential hotspot area set to generate a comprehensive risk feature vector.

[0035] Preferably, the step of using the early warning decision engine to perform hierarchical processing on the comprehensive risk feature vector includes:

[0036] Analyze the temperature anomaly probability and environmental contribution weight in the comprehensive risk feature vector;

[0037] Risk levels are classified according to preset risk thresholds, including low risk, medium risk, and high risk levels;

[0038] Generate an early warning instruction template corresponding to each risk level, wherein the early warning instruction template includes a response action and a parameter range;

[0039] The warning instruction template is dynamically adjusted based on historical warning records to generate an adaptive parameter set;

[0040] Output the final temperature warning command, which integrates the risk level and adaptive parameter set.

[0041] Preferably, the method further includes:

[0042] During the execution of early warning commands, the temperature changes of power transmission lines are monitored in real time, and a temperature feedback data stream is generated.

[0043] Compare the temperature feedback data stream with the expected results of the early warning command, and calculate the deviation index;

[0044] When the deviation index exceeds the tolerance range, the model recalibration process is triggered;

[0045] The parameters of the temperature anomaly detection model are updated based on the deviation index to generate an optimized model version.

[0046] The optimized model version is then reapplied to temperature change feature analysis to iteratively improve the accuracy of early warning.

[0047] Preferably, the triggered model recalibration process includes:

[0048] The temperature feedback data stream is segmented into time windows to generate multiple monitoring time period segments;

[0049] Extract temperature anomalies and their environmental context within each monitoring time segment;

[0050] Calculate the similarity score between the temperature anomaly event and historical patterns to generate pattern matching results;

[0051] The classification boundary of the temperature anomaly detection model is adjusted based on the pattern matching results.

[0052] The model weights are updated using the gradient descent algorithm to generate a recalibrated temperature anomaly detection model.

[0053] Preferably, the method further includes:

[0054] Simulate environmental parameter disturbance scenarios and inject random noise into the multi-source temperature monitoring information;

[0055] Observe the output stability of the early warning decision engine under disturbances and generate stability assessment data;

[0056] When the stability assessment data is lower than the preset benchmark, the reinforcement learning mode of the temperature anomaly detection model is activated;

[0057] The model is trained by using the difference in temperature data before and after the disturbance, which improves its robustness to abnormal patterns.

[0058] Regularly verify model performance to ensure the reliability of early warning commands.

[0059] Preferably, the present invention also includes a temperature warning system based on power transmission lines, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the temperature warning method based on power transmission lines described above.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] This invention significantly improves the comprehensiveness and accuracy of power transmission line temperature monitoring by collecting multi-source temperature monitoring information, encompassing infrared thermal imaging data, contact sensor readings, and environmental parameters. Infrared thermal imaging data can intuitively present the temperature distribution on the conductor surface, allowing maintenance personnel to quickly locate areas with potential temperature anomalies; contact sensor readings can directly obtain temperature values ​​at specific locations on the conductor, ensuring measurement accuracy; the inclusion of environmental parameters such as air temperature, humidity, and wind speed fully considers the impact of the external environment on conductor temperature, enabling the monitoring data to more realistically reflect the actual temperature state of the power transmission line. Compared with traditional single-source monitoring methods, this invention is no longer limited to relying on only one type of data to determine conductor temperature, avoiding information omissions or errors caused by single monitoring methods, thus providing a more reliable and comprehensive data foundation for subsequent analysis and decision-making.

[0062] In processing multi-source temperature monitoring information to extract temperature change characteristics, this invention, through in-depth analysis of spatial temperature gradient distribution and time series trends, can keenly capture abnormal temperature change trends in conductors. Spatial temperature gradient distribution reflects temperature differences between different locations on the conductor; if the temperature gradient in a certain area is significantly abnormal, it likely indicates a potential overheating hazard. Time series trends show the temperature changes over time; by analyzing historical data, the development trend of temperature can be predicted. Combined with a pre-trained temperature anomaly detection model, this invention can accurately generate a set of potential hotspot areas and correlate the probability of temperature exceeding limits with geographical coordinates. This allows maintenance personnel to know in advance which areas are at risk of abnormal temperature increases, as well as the likelihood and specific location of these risks, enabling targeted measures to be taken before a fault occurs. This effectively prevents various accidents caused by excessively high conductor temperatures, ensuring the safe and stable operation of transmission lines.

[0063] This invention integrates a set of potential hotspot areas with real-time environmental parameters to perform multimodal data collaborative assessment, generating a comprehensive risk feature vector. This process makes the assessment of transmission line operational risks more scientific and comprehensive. Different types of data contain information at different levels, and single data points are insufficient to fully reflect the complex operating conditions of transmission lines. Through multimodal data collaborative assessment, the inherent connections between various types of data can be fully explored, allowing for a comprehensive consideration of risks from multiple perspectives. For example, when the probability of temperature exceeding the standard is high in a potential hotspot area, combining real-time environmental parameters such as high temperatures and low wind speeds (conditions unfavorable for heat dissipation) can further determine the degree of risk of a serious fault in that area. This comprehensive assessment method avoids judging risks based solely on a single factor, providing maintenance personnel with more accurate and comprehensive risk information, helping them make more rational decisions and take more effective preventative measures. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the temperature early warning method based on power transmission lines described in this invention.

[0065] Figure 2 A flowchart for multi-source temperature monitoring information acquisition;

[0066] Figure 3 A flowchart for processing multi-source temperature monitoring information and extracting temperature change features;

[0067] Figure 4 This is a comprehensive risk assessment diagram;

[0068] Figure 5 This is a map showing the distribution of early warning levels. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a temperature early warning method and system based on power transmission lines. The method includes: collecting multi-source temperature monitoring information of the power transmission lines, including infrared thermal imaging data, contact sensor readings, and environmental parameters; processing the multi-source temperature monitoring information to extract temperature change features, including spatial temperature gradient distribution and time series trends; analyzing the temperature change features using a pre-trained temperature anomaly detection model to generate a set of potential hotspot areas, which is associated with the probability of temperature exceeding the standard and geographical coordinates; fusing the set of potential hotspot areas with real-time environmental parameters to perform multi-modal data collaborative evaluation and generate a comprehensive risk feature vector; and using an early warning decision engine to perform hierarchical processing on the comprehensive risk feature vector and output a temperature early warning command, which includes the warning level and dynamic adjustment parameters.

[0071] Example 1: See Figure 2The acquisition of multi-source temperature monitoring information for power transmission lines involves continuous infrared scanning of the lines to obtain infrared thermal images and their corresponding timestamp sequences. Infrared scanning is achieved through a network of high-resolution infrared thermal imagers mounted on towers along the transmission corridor. Each imager is equipped with an automatic pan-tilt control system, performing a full-range scan of the transmission line according to a preset scanning path. The scanning frequency is set to complete one full-line scan per minute. Each frame of the infrared thermal image embeds timestamp information accurate to the millisecond level, forming a complete timestamp sequence. The timestamp sequence is synchronized with the Global Positioning System (GPS) clock, providing a unified time reference for subsequent multi-source data fusion. During infrared thermal image acquisition, a temperature calibration algorithm is used to convert the grayscale values ​​of the thermal image into actual temperature values. The calibration algorithm is based on the blackbody radiation law, considering the influence of atmospheric transmittance and environmental radiation, and establishing a mapping relationship between pixel values ​​and actual temperature. Simultaneously, data from contact sensors deployed on the surface of the transmission line is read, generating a sensor reading stream. The contact sensors are platinum resistance temperature sensors, fixed to the surface of the transmission line in a distributed manner with equal spacing. The sensor housing is encapsulated with weather-resistant material, and the interior is filled with thermally conductive silicone grease to ensure temperature response sensitivity. The sensor network is constructed using a wireless sensor network protocol. Each sensor node is equipped with a self-organizing network module to achieve multi-hop data transmission. Sensor reading streams are continuously uploaded at a fixed sampling frequency. Data packets include the sensor number, temperature measurement value, acquisition timestamp, and signal strength indication. The sensor network deployment considers changes in conductor sag and the effects of electromagnetic interference, employing shielded twisted-pair transmission and digital filtering technology to ensure data integrity. Sensor reading streams undergo preliminary verification at the aggregation node, eliminating outliers that significantly exceed the physical measurement range and marking data loss caused by short-term communication interruptions.

[0072] Environmental parameter information collected includes ambient temperature, humidity, wind speed, and solar radiation intensity. One environmental monitoring station is deployed every kilometer along the power transmission line. Each station is equipped with an array of four meteorological sensors. Ambient temperature measurement uses a platinum resistance sensor protected by a radiation shield; humidity measurement uses a capacitive polymer film sensor; wind speed measurement uses an ultrasonic anemometer; and solar radiation intensity measurement uses a thermopile total radiation meter. Environmental parameter information is transmitted to the data center via fiber optic composite overhead ground wire, using the industrial-grade Modbus RTU standard. The data acquisition frequency is synchronized with the infrared scanning. Before being stored, environmental parameter information undergoes quality control system inspection. Inspection rules include range rationality checks, internal consistency checks, and spatiotemporal continuity checks. Detailed logs are recorded for any abnormal data that does not meet quality control requirements. Infrared thermal images, sensor reading streams, and environmental parameter information are time-aligned to generate a time-synchronized multi-source data set. Time alignment uses a distributed clock synchronization scheme based on a network time protocol. All data acquisition devices are connected to the same time server cluster. The time alignment algorithm uses the timestamp of the infrared thermal image as a reference and employs an interpolation algorithm to align the sensor reading streams and environmental parameter information to the same time grid. The interpolation algorithm employs cubic spline interpolation to resample non-uniformly sampled sensor data, generating a data sequence that perfectly matches the timestamps of the infrared thermal image. The time-synchronized multi-source data set is organized and stored in time slices, with each time slice containing all monitoring data collected at the same time. The data storage format uses a hierarchical data model: a top-level time index layer, a middle layer of data type classification, and a bottom layer of raw data values.

[0073] Noise filtering and missing value imputation were performed on multi-source datasets to generate a standardized monitoring dataset. Noise filtering employed differentiated filtering strategies based on the characteristics of different data types. For infrared thermal images, a non-local mean denoising algorithm was used, utilizing weighted averaging based on information from similar pixel blocks in the image to remove random noise while preserving edge details. For contact sensor reading streams, a Kalman filter algorithm was used to establish a state-space model of sensor temperature changes, suppressing measurement noise through a prediction-correction mechanism. For environmental parameter information, a sliding window midpoint filter was used, effectively suppressing sudden interference pulses. Missing value imputation was based on data correlation analysis. Missing pixels in the infrared thermal image were imputed using neighbor pixel weighted interpolation, missing sensor data were imputed using time-series prediction, and missing environmental parameters were imputed using spatial interpolation. Standardization transformed various monitoring data into a unified numerical range. Temperature data standardization used a min-max normalization method, mapping actual temperature values ​​to the [0,1] interval. Environmental parameter standardization used a Z-score standardization method to eliminate the influence of dimensions. The standardized monitoring dataset is organized in matrix form, with the row dimension corresponding to the time series and the column dimension corresponding to the feature variables of different monitoring sources, forming a regular data structure for subsequent feature extraction.

[0074] The acquisition of infrared thermal images must consider the impact of weather conditions. In adverse weather conditions such as rain, snow, and fog, the signal-to-noise ratio of infrared thermal images will significantly decrease. The system is equipped with a weather-adaptive acquisition mode, dynamically adjusting the integration time and gain parameters of the thermal imager based on real-time meteorological data. Reading data from contact sensors requires overcoming strong electromagnetic interference. Sensor nodes employ electromagnetic shielding design and differential signal transmission technology, and the data transmission protocol incorporates a cyclic redundancy check error detection mechanism. The collection of environmental parameter information must ensure spatial representativeness. The location of environmental monitoring stations avoids the influence of local heat sources and obstacles, and the sensor installation height is consistent with the average height of the conductor. The accuracy of time alignment processing directly affects the multi-source data fusion effect. The system uses a combination of hardware and software timestamps. Key data acquisition equipment is equipped with a GPS timing module, and ordinary sensor nodes are synchronized via a network time protocol. Setting noise filtering parameters requires balancing denoising effect and signal fidelity. The optimal parameter combination for each filtering algorithm is determined through historical data testing, and a parameter configuration knowledge base is established to automatically select the filtering scheme based on real-time data characteristics. The accuracy of missing value imputation depends on the identification of missing data patterns. The system constructs a missing value pattern classifier and selects the most suitable imputation algorithm based on features such as the duration and distribution of missing values. The generation of standardized monitoring datasets needs to maintain the data distribution characteristics. The standardization parameters are calculated based on the statistical characteristics of long-term monitoring data and are updated periodically to adapt to seasonal changes.

[0075] The acquisition of multi-source temperature monitoring information is a systematic project involving the collaborative work of various measurement technologies, including optical measurement, contact measurement, and environmental measurement. Clock synchronization between measuring devices is achieved through a precise time protocol, with synchronization accuracy reaching the microsecond level. In addition to temperature information, infrared thermal image acquisition also records the spatial attitude parameters of the devices for subsequent image geometric correction. Data transmission from contact sensors employs a time-division multiple access mechanism to avoid wireless channel conflicts, and sensor nodes are equipped with solar power systems to ensure long-term continuous operation. Environmental parameter information collection includes quality control metadata, with each data point accompanied by a quality flag indicating the data's reliability level. Time alignment processing uses a sliding window mechanism to buffer data packets with large transmission delays, ensuring the correctness of data timing. Noise filtering is implemented layered between edge computing nodes and cloud servers; simple filtering is performed at the edge, while complex filtering is performed in the cloud. The missing value imputation algorithm considers the randomness and systematic differences in data missingness, employing probabilistic imputation methods for random missing values ​​and causal inference imputation for systematic missing values.

[0076] Example 2: See Figure 3The process of processing multi-source temperature monitoring information to extract temperature change characteristics is based on a standardized monitoring dataset. This dataset comes from the regularized data matrix generated in Example 1. Spatial grid partitioning employs a uniform segmentation strategy under a geographic coordinate system, discretizing the continuous geographic area covered by the power transmission line into a regular array of grid cells. The size of each grid cell is determined based on the spatial resolution of the infrared thermal image, and the grid side length is set to an integer multiple of the ground size corresponding to the instantaneous field of view of the infrared thermal imager, ensuring that each grid cell contains a sufficient number of infrared pixels for statistical analysis. Grid partitioning considers the actual route of the power transmission line and topographic relief, and uses a digital elevation model to perform elevation correction on the grid, eliminating spatial distortion caused by topographic factors. Each grid cell is assigned a unique identifier, and the identifier encoding rule includes region code, row number, and column number information, establishing a precise mapping relationship between the grid cell and geographic coordinates. The spatial grid partitioning results in multiple temperature analysis regions, each corresponding to an independent temperature analysis cell. The temperature analysis cell contains temperature data of all infrared pixels within the grid, contact sensor readings, and environmental parameters. The magnitude and direction of the temperature gradient within each temperature analysis region are calculated using a differential geometry method. The temperature gradient calculation is based on the temperature distribution within and around the grid cells, constructing a two-dimensional discrete function of the temperature field. The gradient magnitude is calculated using the central difference algorithm, calculating the first-order partial derivatives in the east-west and north-south directions of the grid cell. The gradient magnitude is the Euclidean norm of the partial derivatives in both directions. The gradient direction is calculated as the azimuth angle of the partial derivative vector, representing the direction of the fastest temperature change. The spatial gradient distribution map is generated using vector field visualization technology. A gradient arrow is drawn at the center point of each grid cell, with the arrow length representing the gradient magnitude and the arrow pointing to the gradient direction. The spatial gradient distribution map is stored in raster data format, with each pixel containing two floating-point values: gradient magnitude and direction. Geometric boundary metadata of the grid cells is also stored. A boundary handling strategy is employed during the temperature gradient calculation process. Grid cells at the region boundaries are expanded using a mirror-fill method to ensure the integrity of the gradient calculation across the entire region.

[0077] The extraction of time series trend features is performed on the temperature data series for each temperature analysis region. The temperature data series comes from temperature observations at consecutive timestamps in a standardized monitoring dataset. The temperature change rate feature is obtained through first-order differencing, calculating the difference ratio between temperature values ​​at adjacent time points to reflect the instantaneous rate of temperature change. Periodic fluctuation features are separated using a seasonal-trend decomposition method, employing the LOESS locally weighted regression algorithm to extract the long-term trend term, seasonal term, and residual term of the temperature series. The periodic fluctuation feature primarily extracts amplitude and phase parameters from the seasonal term. Anomaly fluctuation pattern detection utilizes an analysis framework based on prediction bias. Temperature prediction values ​​are generated using the ARIMA time series prediction model, and significant deviations between actual and predicted values ​​are identified as anomalies. The time series trend features are represented as multi-dimensional feature vectors, with one feature vector corresponding to each temperature analysis region. Vector dimensions include multiple indicators such as short-term change rate, long-term trend slope, periodic amplitude, periodic phase, and anomaly fluctuation intensity. The time span of the feature vector is dynamically updated using a sliding window mechanism, with the window size adaptively adjusted according to the characteristic time scale of temperature change. The tensor concatenation operation between the spatial gradient distribution map and the time series trend features is performed in the feature space. The spatial gradient distribution map is represented as a second-order tensor structure, with the two dimensions of the tensor corresponding to the row and column indices of the grid. Each grid position contains two feature channels: gradient magnitude and direction. The time series trend features are expanded into first-order feature vectors, with a vector length equal to the number of feature dimensions. The tensor concatenation operation is performed along the channel dimension, expanding the time series feature vectors into a multi-channel tensor with the same spatial dimension as the spatial gradient distribution map through a broadcast mechanism. The number of channels in the new tensor is the sum of the original gradient channel number and the time feature dimension. The multi-dimensional temperature change feature matrix is ​​generated after dimension alignment processing, and the different scales of the spatial gradient and time features are unified to the same numerical range through max-min normalization. The multi-dimensional temperature change feature matrix is ​​stored in a three-dimensional array format, with the three dimensions corresponding to the spatial row coordinates, spatial column coordinates, and feature channel indices, respectively. The matrix elements are in single-precision floating-point format to balance storage efficiency and computational precision.

[0078] The process of analyzing temperature change features using a pre-trained temperature anomaly detection model takes a multi-dimensional temperature change feature matrix as input. The convolutional neural network (CNN) architecture employs an encoder-decoder structure, with the encoder containing multiple convolutional and pooling layers. The convolutional layers use small 3×3 kernels for local feature extraction, with a stride of 1 pixel and zero-padding to maintain the spatial size of the feature map. Each convolutional layer is followed by a ReLU activation function to introduce a non-linear transformation, and the pooling layers use 2×2 max pooling to downsample the feature map. The CNN is trained based on a historical temperature anomaly case dataset, using a cross-entropy loss function and the Adam optimizer to iteratively update the network weight parameters. The feature activation map is generated from the output of the last convolutional layer of the CNN, preserving the spatial structure information of the input matrix. Each activation map channel corresponds to a temperature anomaly pattern response. A region proposal network (RPN) processes the feature activation map, generating multiple candidate bounding boxes at each spatial location. The RPN consists of two parallel branches: one predicts the bounding box's positional offset, and the other predicts the confidence score that the bounding box contains the anomaly region. The generation of candidate anomaly regions employs a multi-scale anchor box mechanism. Preset anchor box templates with different aspect ratios are slidably scanned across the feature activation map, predicting nine candidate regions at each anchor box location. The confidence score is normalized to a probability value using a softmax function, representing the likelihood that the candidate region contains a temperature anomaly. The region proposal network is trained end-to-end, sharing feature extraction layer parameters with the convolutional neural network. A multi-task loss function simultaneously optimizes the accuracy of region classification and bounding box regression.

[0079] The process of using the Non-Maximum Suppression (NMS) algorithm to screen candidate outlier regions is based on confidence scores and overlap metrics. The NMS algorithm iterates through all candidate regions, sorting them by confidence score from highest to lowest, and sequentially selecting the candidate region with the highest score as the retained object, while suppressing other candidate regions with high overlap. The overlap is calculated using the Cross-Union Ratio (CUI) metric, with a CUI threshold set to 0.7; candidate regions exceeding this threshold are considered redundant detections. The NMS algorithm uses a greedy strategy for iterative processing until all candidate regions have been processed, generating a deduplicated candidate region set. During the screening process, the spatial coordinates and confidence scores of each candidate region are retained, establishing a correspondence between candidate regions and the original feature maps. The step of generating a potential hotspot region set based on the confidence threshold sets a dynamic confidence threshold. The confidence threshold is determined based on the precision-recall curve of the model on the validation set, with the threshold that maximizes the F1 score chosen as the default value. The potential hotspot region set is generated by filtering low-confidence candidate regions, retaining only those with confidence scores higher than the threshold. Each potential hotspot region contains two core attributes: center coordinates and temperature anomaly probability. The center coordinates are calculated through bounding box coordinate transformation, while the temperature anomaly probability is directly derived from the confidence score of the region proposal network. The set of potential hotspot regions is organized in a structured data format, with each region instance containing complete metadata such as a region identifier, center point latitude and longitude coordinates, bounding box size, temperature anomaly probability value, and timestamp. The set of potential hotspot regions is sorted from high to low temperature anomaly probability, facilitating subsequent processing by prioritizing high-probability anomaly regions.

[0080] The pre-training process of the temperature anomaly detection model utilizes a large-scale labeled temperature dataset containing samples of various typical anomaly patterns. Data augmentation strategies are employed during model training to enhance generalization ability. The feature extraction capability of the convolutional neural network is achieved through a deep network structure, with the network depth chosen based on a balance between computational resources and the usage scenario. The multi-scale anchor box design of the region proposal network adapts to temperature anomaly regions of varying sizes, effectively detecting everything from local hotspots to distributed temperature anomalies. The parameter configuration of the non-maximum suppression algorithm affects the recall and precision of the detection results; parameter tuning is based on validation set performance evaluation. The confidence threshold setting considers the balance between false positive and false negative rates and can be dynamically adjusted according to risk preferences in practical applications. The spatial distribution of the potential hotspot region set reflects the temperature anomaly situation of the transmission lines, providing fundamental data support for subsequent risk assessment. The generation of the multi-dimensional temperature change feature matrix integrates both spatial gradient and temporal series feature modes. This multi-modal feature representation can simultaneously capture the spatial distribution characteristics and temporal evolution patterns of temperature anomalies. The local receptive field design of the convolutional neural network makes it sensitive to local spatial patterns, suitable for detecting regional features of temperature anomalies. The anchor box mechanism of the region proposal network draws on advanced techniques in object detection, transforming temperature anomaly region detection into a bounding box regression problem. Non-maximum suppression algorithms address the deduplication of multiple overlapping candidate regions, ensuring that each anomaly region corresponds to only one detection result. Confidence threshold filtering reduces the number of false alarms, improving the reliability of the early warning system. The spatial coordinate information of potential hotspot region sets is integrated with a geographic information system, supporting the visual location display of temperature anomalies.

[0081] Example 3: The operation of fusing potential hotspot area sets with real-time environmental parameters is based on spatial location matching. A subset of environmental parameters for the corresponding location is extracted based on the center coordinates of the potential hotspot area set. The center coordinates are derived from the geolocation information of the hotspot areas generated in Example 2. The extraction of the environmental parameter subset uses a spatial nearest neighbor query algorithm. Using the center point of each hotspot area as the center, a search radius threshold is set, and all monitoring stations falling within the circular area are searched in the environmental monitoring station database. The search radius is determined considering the spatial correlation decay distance of environmental parameters. Based on historical data analysis of the spatial variogram model of environmental parameters, 80% of the range distance is used as the default search radius. The environmental parameter subset contains environmental data from multiple monitoring stations. Data from multiple stations corresponding to the same hotspot area are fused using an inverse distance weighted average method to generate a comprehensive parameter value representing the environmental condition of that location. The timestamps of the environmental parameter subset are strictly aligned with the detection time of the potential hotspot area set to ensure temporal consistency between environmental data and temperature anomaly detection results. The feature encoding conversion process of the environmental parameter subset maps multi-dimensional environmental parameters into fixed-dimensional environmental feature vectors. The environmental parameters include four physical quantities: ambient temperature, humidity, wind speed, and solar radiation intensity. Feature encoding employs a field-based processing strategy. Environmental temperature parameters are transformed to the zero-to-one interval using a minimum-maximum normalization method. Humidity parameters are directly used as feature values ​​in percentage form. Wind speed parameters undergo logarithmic transformation to address skewed distributions, and solar radiation intensity is compressed using a square root transformation. The dimensional design of the environmental feature vector considers the interactions between parameters, adding derivative features such as interaction terms between humidity and temperature, and wind speed and radiation, in addition to the four basic parameters. Each dimension of the environmental feature vector is assigned a feature name with clear physical meaning, and a correspondence table between feature indexes and parameter types is established. The unit of measurement information of the original parameters is preserved during feature encoding, and the unit conversion coefficient for each feature dimension is recorded in the feature vector metadata.

[0082] The calculation of the weighted fusion of environmental feature vectors and the temperature anomaly probability of potential hotspot area sets adopts a linear weighted model, and the weighted fusion formula is expressed as:

[0083] in: Indicates the integration risk score, Indicates the probability of temperature anomalies. The first element representing the environmental feature vector One dimension, Weighting coefficients representing the probability of temperature anomalies. Indicates the first The weighting coefficients of each environmental feature. This represents the total number of dimensions in the environmental feature vector. Weight coefficients are determined through feature importance analysis. The random forest algorithm is used to calculate the contribution of each feature to temperature anomaly prediction, and the normalized contribution value is used as the initial weight. Before weighted fusion calculation, the environmental feature vector is standardized to ensure that the feature values ​​of each dimension have the same scale range, avoiding interference from differences in numerical magnitude on weight allocation.

[0084] The operation of calculating the contribution weights of environmental factors to temperature anomalies using an attention mechanism is based on a query-key-value pair mechanism. The query vector of the attention mechanism is derived from the temperature anomaly probability, the key vector is composed of environmental feature vectors, and the value vector is the same as the key vector. The attention weights are calculated using a scaled dot product attention formula. First, the dot product between the query vector and each key vector is calculated, and then normalized to an attention distribution using a softmax function. The attention weights reflect the relative importance of different environmental factors to the current temperature anomaly pattern; a high weight indicates a strong correlation between the environmental factor and the temperature anomaly. The dynamic nature of the attention weights allows them to adapt to temperature anomaly mechanisms under different meteorological conditions. In dry, windy weather, the weight of wind speed may increase, and in hot, humid weather, the weight of humidity may increase. The calculation results of the attention mechanism generate a set of weight vectors with the same number of dimensions as the environmental features. The weight vectors are multiplied element-wise with the original environmental feature vectors to produce weighted environmental features. The operation of associating the weighted fusion results with the geographic coordinates of a set of potential hotspot areas constructs a spatially explicit comprehensive risk feature vector. The association mapping uses a spatial connection method, attaching the weighted fusion results as attribute data to the spatial location of the potential hotspot areas. The data structure design of the comprehensive risk feature vector includes spatial and attribute fields. The spatial fields record the geometric information of hotspot areas, including center point coordinates, bounding box range, area, and other spatial attributes. The attribute fields contain risk assessment parameters such as temperature anomaly probability, environmental feature weighting, fused risk score, and attention weight distribution. The generation of the comprehensive risk feature vector considers the temporal dimension; each feature vector is accompanied by a timestamp to record the effective time point of the risk assessment. The feature vectors are stored using a spatiotemporal data model, supporting efficient querying and retrieval by time series and spatial range.

[0085] The extraction accuracy of environmental parameter subsets directly affects the fusion effect. The spatial distribution density of environmental monitoring stations needs to satisfy the Nyquist sampling theorem to avoid estimation bias caused by spatial aliasing. The rationality of feature encoding affects the representation ability of environmental factors. The encoding scheme needs to retain the physical meaning of environmental parameters while eliminating the influence of dimensions. The weight coefficients of the weighted fusion formula need to be updated regularly to adapt to parameter drift caused by seasonal changes and equipment aging. The computational complexity of the attention mechanism is quadratically related to the number of environmental feature dimensions. In practical applications, a trade-off between model complexity and computational efficiency is required. The spatial accuracy of the association mapping depends on the uniformity of the coordinate system. All spatial data needs to be transformed to the same map projection coordinate system. The time lag effect of environmental parameters needs to be considered in feature encoding. The impact of some environmental factors on temperature anomalies has a time delay. Feature encoding introduces time-delayed environmental parameters as supplementary features. The linear assumption of the weighted fusion formula may not be sufficient in practical applications. Nonlinear fusion terms can be introduced to enhance the model's expressive power. The interpretability of the attention mechanism makes it suitable for application in safety-critical systems. The attention weight distribution can serve as auxiliary explanatory information for decision-making. The dimensionality of the comprehensive risk feature vector may be high, necessitating the use of feature selection methods to remove redundant dimensions and improve subsequent processing efficiency. The spatial interpolation accuracy of environmental parameters affects the accuracy of risk assessment in remote areas, requiring the use of advanced spatial interpolation algorithms such as Kriging interpolation to improve estimation accuracy.

[0086] The spatial distribution patterns of potential hotspot areas are noteworthy in association mapping; clustered hotspots may indicate systemic risks, while dispersed distributions may reflect localized problems. Strict standards are needed for quality control of environmental parameter subsets, and statistical testing methods should be used to identify and remove anomalous environmental data. The robustness of feature encoding is enhanced through outlier handling techniques, and winsorizing is used to shrink extreme environmental parameter values. The weight coefficients of the weighted fusion formula can be dynamically adjusted through an online learning mechanism, optimizing weight allocation based on feedback signals from early warning results. The calculation of the attention mechanism can incorporate location encoding information, enhancing the model's understanding of spatial distribution patterns. The spatial analysis function of association mapping can integrate GIS technologies such as buffer analysis and overlay analysis to deepen the spatial dimension insight of risk assessment. The generation of the comprehensive risk feature vector marks the completion of multimodal data fusion; the temperature anomaly information and environmental context information contained in the vector provide comprehensive input for early warning decisions. The standardized format of the feature vector facilitates subsequent processing by machine learning models, and the unified interface specification supports modular system design. The quantifiable nature of risk assessment parameters makes risk results from different regions and times comparable, supporting risk trend analysis and historical comparison.

[0087] See Figure 4This chart visually presents the risk levels of different regions through a geospatial distribution. Each data point represents a potential hotspot area, with its location corresponding to actual geographic coordinates for easy spatial positioning analysis. The chart uses color gradients to represent the overall risk score, with a color change from blue to red indicating an increasing risk level. This visual coding method allows for the rapid identification of high-risk areas. Simultaneously, the size of the data points reflects the coverage density of environmental monitoring stations in the surrounding area; larger points indicate more monitoring data supporting the assessment, making the results more reliable. Through this chart, maintenance personnel can clearly understand the temperature risk distribution pattern of the entire transmission line and identify areas requiring focused attention. This spatially visualized risk assessment method provides an important basis for developing targeted inspection plans and preventative measures, contributing to a shift from a reactive to a proactive maintenance model.

[0088] Example 4: The process of hierarchically processing the comprehensive risk feature vector using the early warning decision engine begins with deep parsing of the vector. Parsing the temperature anomaly probability and environmental contribution weights in the comprehensive risk feature vector requires extracting key fields from the vector's data structure. The comprehensive risk feature vector adopts a hierarchical storage structure. The top layer contains metadata information such as timestamps and regional identifiers, while the data layer contains numerical fields such as temperature anomaly probability values, environmental feature weights, and fused risk scores. The parsing algorithm locates the target data through a field mapping table. The temperature anomaly probability is stored in single-precision floating-point format from the sixth to the ninth byte of the vector, and the environmental contribution weights are stored in an array structure starting from the twelfth byte. The parsing process verifies data validity, checking whether the numerical range is within a reasonable range. For outliers exceeding physical meaning, a data cleaning process is initiated. The parsing of the environmental contribution weights needs to consider the normalization of the weight distribution, verifying whether the sum of all weight values ​​is close to one unit. Weight sets that do not meet the normalization condition are re-standardized. The operation of classifying risk levels according to preset risk thresholds depends on the configuration of the threshold parameter table. Risk levels are divided into three main levels: low risk, medium risk, and high risk, each corresponding to a different early warning response strategy. The risk threshold parameter table is stored in an enterprise-level database. Threshold values ​​are determined based on statistical analysis of historical accident data and expert experience, and are dynamically adjusted with seasonal changes. Risk level classification employs a multi-dimensional decision boundary, with the probability of temperature anomalies setting the primary threshold, and environmental contribution weight serving as a moderating factor to adjust the level boundary. The criteria for low-risk levels are a fusion risk score below 0.3, medium-risk levels correspond to a score range of 0.3 to 0.7, and high-risk levels require a score exceeding 0.7. Fuzzy logic processing is implemented for risk level classification, setting transition regions near the threshold boundaries to avoid decision-making fluctuations caused by abrupt changes in level. The risk level classification results include a confidence assessment, calculated based on the degree of matching between current data and historical patterns.

[0089] The generation of early warning instruction templates corresponding to each risk level is based on a template library. These templates are defined in XML format and include structured information such as a list of response actions, parameter value ranges, and execution conditions. Low-risk early warning instruction templates specify an increase in routine inspection frequency to once every two hours, a shortened load monitoring cycle to ten minutes, and internal system message notification as the alarm method. Medium-risk early warning instruction templates require an hourly on-site inspection frequency, a load adjustment range limited to within 10%, and SMS notification to maintenance personnel as the alarm method. High-risk early warning instruction templates initiate an emergency response process, increase the inspection frequency to real-time continuous monitoring, allow a 30% load reduction, and trigger audible and visual alarms and multiple telephone notifications as alarm methods. The parameter range settings for the early warning instruction templates consider equipment safety operation limitations, with parameter value upper and lower bounds dynamically calculated based on conductor type and environmental conditions. The early warning instruction template generation mechanism supports version management; different versions are suitable for different types of transmission lines and weather conditions. Dynamic adjustment of the early warning instruction templates based on historical early warning records enables adaptive parameter optimization. The historical early warning record database stores full-process data for each early warning event, including early warning level, response actions, handling results, and effect evaluation. The dynamic adjustment algorithm analyzes historical success patterns and failure cases to establish a correlation model between early warning parameters and response effectiveness. The adaptive parameter set is generated using machine learning algorithms, employing a decision tree model to analyze the mapping relationship between parameter settings and response effectiveness, and optimizing parameter combinations. A feedback control mechanism is introduced into the dynamic adjustment process, using recent early warning response effects as adjustment signals and employing a PID controller principle to progressively optimize parameter values. The adaptive parameter set is updated every 24 hours to ensure that parameter adjustments promptly reflect changes in the system's operating status. Limits are set on the magnitude of parameter changes during dynamic adjustment to prevent excessively large single adjustments from causing system oscillations.

[0090] The process of outputting the final temperature warning command integrates the aforementioned processing results into a standardized message. The temperature warning command is encapsulated in JSON data format, with the message header containing management information such as the command number, generation time, and validity period. The message body integrates a risk level code and an adaptive parameter set. The risk level code is represented by a three-digit code, with the first digit indicating the risk category and the last two digits indicating the risk intensity. The adaptive parameter set is organized in key-value pairs, with each parameter containing a parameter name, parameter value, unit of measurement, and applicable scope marker. The output interface for the temperature warning command supports multiple communication protocols, including HTTP RESTful API, MQTT message queues, and TCP socket transmission. A digital signature is generated synchronously during command output, and RSA asymmetric encryption is used to ensure the integrity and authenticity of the command. The transmission priority of the temperature warning command is set according to the risk level, with high-risk commands enjoying the highest transmission priority to ensure timely delivery of warning information. The rule base of the warning decision engine needs to be updated and maintained regularly. Rule updates are based on the latest accident statistics and equipment operation data, with a comprehensive monthly evaluation cycle. The adjustment of the risk threshold parameter table takes into account seasonal variations; summer thresholds are typically more stringent than winter thresholds because download speeds decrease in high-temperature environments. The version control of the early warning command template adopts a canary release mechanism. New versions are first piloted on select lines to verify their effectiveness before being rolled out nationwide. Analysis of historical early warning records utilizes data mining techniques to identify potential risk patterns and optimization opportunities. The optimization of the adaptive parameter set balances security and cost-effectiveness, minimizing unnecessary load adjustments while ensuring safety. The reliability of temperature early warning command transmission is guaranteed through a retransmission mechanism, and important commands require confirmation acknowledgments from the receiver.

[0091] Referring to Table 1, the accuracy of risk level classification directly affects the early warning effect. The classification algorithm needs to be backtested regularly using historical data to evaluate the accuracy and recall rate of the classification. Personalized settings for early warning command templates consider regional differences; different parameter ranges can be configured for mountainous and plain routes. Dynamically adjusting the learning efficiency of the algorithm is a crucial indicator; the algorithm needs to quickly extract effective experience from limited historical data. Standardized formats for temperature early warning commands facilitate integration with other systems, such as power grid dispatching systems and equipment management systems. The computational efficiency of the early warning decision engine needs to meet real-time requirements, ensuring timely output of early warning commands even during peak data periods. The visualization of risk level classification results uses color coding: low risk is represented by blue, medium risk by yellow, and high risk by red. The field design of the early warning command template considers scalability, reserving spare fields for future functional expansion. The data structure of historical early warning records includes time-series information, supporting trend analysis and pattern recognition. Safety boundaries are set for the adjustment range of the adaptive parameter set to prevent parameter adjustments from exceeding reasonable limits. The transmission encryption of temperature early warning commands uses national cryptographic algorithms to ensure data transmission security. The redundant design of the early warning decision engine improves system reliability, and hot switching is achieved between the primary and backup engines.

[0092] Table 1: Risk Level Threshold Classification Standards

[0093]

[0094] The criteria for risk level classification need to be dynamically adjusted based on actual conditions. The threshold ranges in the table represent baseline values, which can be fine-tuned in practice based on line importance and environmental conditions. The integrated risk scoring range uses a left-closed, right-open interval division to ensure that each score belongs to only one risk level. The temperature anomaly probability threshold serves as the primary criterion, while an environmental weighting adjustment coefficient is used to correct the confidence level of the probability value. Inspection frequency requirements increase progressively with risk levels, with higher risk levels requiring more intensive monitoring. Load adjustment limits consider the needs of stable grid operation, minimizing power supply impact within a controllable risk range.

[0095] See Figure 5This chart demonstrates the distribution of early warning levels after the early warning decision engine has processed the comprehensive risk feature vector into different levels. A pie chart clearly shows the proportion of different risk levels in the overall assessment, facilitating a macro-level understanding of the overall risk situation. The chart divides the early warning levels into three main categories according to risk severity, each with a corresponding color code and detailed operational parameters. Low-risk areas are represented in green, corresponding to more lenient monitoring requirements; medium-risk areas are highlighted in orange, requiring enhanced monitoring; and high-risk areas are highlighted in red, demanding immediate emergency response measures. The legend details the specific early warning parameters for each risk level, including key indicators such as inspection frequency requirements and load adjustment limits. This tiered early warning mechanism ensures the rational allocation of limited operational resources, guaranteeing timely handling of high-risk areas while avoiding excessive resource allocation in low-risk areas, achieving a good balance between safety and economy.

[0096] Example 5: Real-time monitoring of transmission line temperature changes during the execution of early warning commands forms the basis of the system's closed-loop regulation. The generation of the temperature feedback data stream relies on the continuous operation of a sensor network deployed on the transmission lines. Taking the operation of a 500 kV transmission line during the high-temperature period in summer as an example, the temperature feedback data stream contains continuous readings from twelve temperature measurement points. Each measurement point is equipped with three redundant sensors, and the data sampling interval is set to ten seconds. The temperature feedback data stream is transmitted to the monitoring center via an optical fiber composite overhead ground wire. The transmission protocol adopts the IEC61850 standard, and the data packet includes fields such as sensor ID, temperature value, timestamp, and quality code. When the temperature feedback data stream is accessed by the monitoring system, it undergoes a data quality check, eliminating outliers that are significantly beyond physical possibility. Linear interpolation is used to complete data missing due to short-term communication interruptions. The temperature feedback data stream is stored using a time-series database structure, supporting efficient time range queries and streaming processing. The analysis comparing the temperature feedback data stream with the expected results of the early warning commands requires establishing a benchmark. The expected results of the early warning commands are derived from the target temperature range set in the temperature early warning commands output in Example 4. In a specific case, the system issued a medium-risk warning for conductor segment TL-2034, with the expected result being that the temperature in that segment be controlled below 60 degrees Celsius. The comparison operation employs a sliding window mechanism, calculating the average actual temperature within a five-minute time window and comparing it with the expected value of the warning command. The deviation index is calculated using a multi-dimensional evaluation method, including three indicators: mean absolute error, root mean square error, and maximum deviation value. The mean absolute error reflects the overall deviation level, the root mean square error is more sensitive to large deviations, and the maximum deviation value identifies the most severe deviation. The deviation indicators are summarized every thirty minutes, generating a deviation report for decision-making purposes.

[0097] The judgment when the deviation index exceeds the tolerance range is based on a preset tolerance threshold. The tolerance range is dynamically set according to the wire type and safety standards. In the example case, the tolerance range is set to ±5 degrees Celsius. The trigger condition for the deviation index exceeding the tolerance range is that it exceeds the threshold for three consecutive monitoring cycles to avoid false triggers caused by accidental fluctuations. The instruction to trigger the model recalibration process is automatically generated by the monitoring system. The trigger signal includes contextual information such as deviation details, time range, and affected area. The start of the model recalibration process requires confirmation from the system administrator. After confirmation, the system enters calibration mode, suspends the automatic update of warning instructions, and maintains the current instruction to continue execution. The operation of updating the parameters of the temperature anomaly detection model based on the deviation index adopts an incremental learning strategy. The parameters of the temperature anomaly detection model are stored in the network weight matrix. In specific implementation, the model recalibration uses the temperature feedback data of the most recent 30 days as training samples. The sample data includes normal temperature patterns and abnormal temperature patterns. Parameter updates are implemented through the gradient descent algorithm. The error gradient between the predicted output and the actual temperature is calculated, and the network weights are adjusted along the opposite direction of the gradient. The weight update step size adopts an adaptive learning rate algorithm. The initial learning rate is set to 0.001 and dynamically adjusted according to the convergence situation. Historical versions are retained during model parameter updates, and a new model version number is generated with each update, which facilitates version rollback and effect comparison.

[0098] The process of reapplying the optimized model version to temperature change feature analysis requires testing and verification. The optimized temperature anomaly detection model is first tested offline on historical data, with the test dataset including various typical scenarios. After passing the tests, the model is deployed to the production environment using a blue-green deployment approach. The new model version is first tested on a small number of lines, with the old and new models working in parallel during the trial run, and the output results are compared. The model switch is performed during periods of low load to minimize the impact on system operation. After the optimized model version is officially launched, its performance indicators are continuously monitored, including warning accuracy, false alarm rate, and response time. The specific operation of triggering the model recalibration process includes several sub-steps. The method for segmenting the temperature feedback data stream into time windows uses a fixed-length overlapping window, with a window length set to 60 minutes and a sliding step size of 10 minutes. In the example case, 24 hours of continuous temperature data is divided into 138 effective window segments, each containing 360 temperature sampling points. A mutation detection algorithm is used to extract temperature anomaly events within each monitoring time segment, identifying peaks and step changes in the temperature curve. The characteristics of temperature anomaly events include parameters such as start time, duration, amplitude change, and rate of change. The environmental context was extracted from the same period's meteorological database, including conditions such as wind speed, sunshine, and humidity at the time of the temperature event.

[0099] The analysis of similarity scores between temperature anomaly events and historical patterns employs a dynamic time warping algorithm, which can handle comparisons of time series of varying lengths. In a specific case, the system detects a 25-minute temperature spike and calculates its similarity with 300 typical patterns stored in the historical database. Similarity scores range from zero to one; higher scores indicate closer similarity to historical patterns, while patterns with scores below 0.3 are classified as novel. The pattern matching output includes the three most similar historical patterns and their similarity scores, providing a reference for model adjustment. Adjusting the classification boundary of the temperature anomaly detection model based on the pattern matching results is performed in the feature space. The classification boundary of the temperature anomaly detection model is a hyperplane distinguishing between normal and abnormal temperatures. For historical patterns with high similarity, the classification boundary remains relatively stable; for novel patterns, the classification boundary needs to be expanded to accommodate new anomaly patterns. The classification boundary adjustment uses the boundary relaxation method from support vector machine theory, appropriately increasing the boundary tolerance while ensuring classification accuracy. The boundary adjustment amount is calculated based on the similarity score; the lower the similarity, the larger the boundary adjustment.

[0100] The process of updating model weights using the gradient descent algorithm requires multiple iterations. A variant of mini-batch stochastic gradient descent is used, with each iteration employing 32 samples to calculate the gradient. The momentum term in the weight update formula is set to 0.9 to reduce oscillations during optimization. A regularization term is added during the model weight update process to prevent overfitting; the regularization coefficient is determined through cross-validation. Testing under simulated environmental parameter disturbances is a crucial step in model validation. The injection of random noise into multi-source temperature monitoring information utilizes a Gaussian white noise model, with the noise amplitude set to 5% of the signal amplitude. In the test case, the system injects different types of noise into infrared thermal imaging data, sensor readings, and environmental parameters. Salt-and-pepper noise is injected into the infrared thermal imaging data to simulate rain and snow interference, Gaussian noise into the sensor readings to simulate measurement errors, and impulse noise into the environmental parameters to simulate transmission anomalies. Disturbance testing is conducted in an isolated environment to avoid impacting the actual production system.

[0101] Monitoring the stability of the early warning decision engine's output under disturbances requires designing evaluation metrics. Stability evaluation data includes statistics such as the variance of the output results, fluctuation range, and number of jumps. In the example test, the system added ten different noise patterns to the same set of temperature data and ran the early warning decision engine to obtain ten sets of output results. The stability evaluation data calculates the consistency of these ten sets of results, including the consistency of risk levels and the magnitude of changes in early warning parameters. The criterion for stability evaluation data falling below a preset benchmark is that the variance of the output results exceeds a threshold or a jump in risk level occurs. When the stability evaluation data falls below the preset benchmark, the response process initiates a reinforcement learning mode. The reinforcement learning mode learns the optimal decision strategy through interaction between the agent and the environment. In the reinforcement learning mode, the temperature anomaly detection model acts as the agent, the temperature monitoring environment acts as the environment, and the reward function is designed based on early warning accuracy and stability. The process of training the model using the difference in temperature data before and after the disturbance constructs a state-action-reward sequence. The state is the temperature feature vector, the action is the adjustment of model parameters, and the reward is the degree of stability improvement. The training process uses the Q-learning algorithm to continuously optimize the decision strategy through trial and error. The mechanism for periodically validating model performance is executed on a fixed schedule, once a week, using actual running data from the most recent week. Validation includes traditional metrics such as model accuracy, recall, and F1 score, as well as performance metrics such as response latency and resource consumption. Validation results are compared with historical performance, and an alert is triggered when a performance decline is detected. The reliability of these alerts is ensured through multiple safeguards, including model performance monitoring, anomaly detection, and automatic recovery.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A temperature early warning method based on power transmission lines, characterized in that, include: Collect multi-source temperature monitoring information of power transmission lines, including infrared thermal image data, contact sensor readings, and environmental parameters; The multi-source temperature monitoring information is processed to extract temperature change features, which include spatial temperature gradient distribution and time series trends. The temperature change characteristics are analyzed by a pre-trained temperature anomaly detection model to generate a set of potential hotspot areas, which are associated with the probability of temperature exceeding the standard and geographical coordinates. By integrating the set of potential hotspot areas with real-time environmental parameters, a multimodal data collaborative assessment is performed to generate a comprehensive risk feature vector; The comprehensive risk feature vector is processed in a hierarchical manner using an early warning decision engine, and a temperature early warning command is output. The temperature early warning command includes the early warning level and dynamic adjustment parameters.

2. The temperature early warning method based on power transmission lines as described in claim 1, characterized in that, The multi-source temperature monitoring information collected from the power transmission lines includes: Continuous infrared scanning of the transmission line was performed to obtain infrared thermal images and their corresponding timestamp sequences; Synchronously read data from contact temperature sensors arranged on the surface of the conductors to generate a sensor reading stream; Collect environmental parameter information, including ambient temperature, humidity, wind speed, and solar radiation intensity; The infrared thermal image, sensor reading stream, and environmental parameter information are time-aligned to generate a time-synchronized multi-source data set. The multi-source data set is subjected to noise filtering and missing value imputation to generate a standardized monitoring dataset.

3. The temperature early warning method based on power transmission lines as described in claim 2, characterized in that, The process of processing the multi-source temperature monitoring information to extract temperature change features includes: The standardized monitoring dataset is divided into spatial grids to generate multiple temperature analysis regions; Calculate the magnitude and direction of the temperature gradient within each temperature analysis region to generate a spatial gradient distribution map; Extract time series trend features, including temperature change rate, periodic fluctuations, and abnormal fluctuation patterns; The spatial gradient distribution map and the time series trend features are tensor-concatenated to generate a multidimensional temperature change feature matrix.

4. The temperature early warning method based on power transmission lines as described in claim 3, characterized in that, The analysis of the temperature change characteristics using a pre-trained temperature anomaly detection model includes: The multidimensional temperature change feature matrix is ​​input into a convolutional neural network for local feature extraction to generate a feature activation map. The feature activation map is processed by a region proposal network to generate candidate anomaly regions and their confidence scores; The nonmaximum suppression algorithm is applied to filter the candidate abnormal regions and remove regions with high overlap. A set of potential hotspot regions is generated based on a confidence threshold. Each potential hotspot region includes its center coordinates and the probability of temperature anomalies.

5. The temperature early warning method based on power transmission lines as described in claim 4, characterized in that, The fusion of the potential hotspot area set and real-time environmental parameters includes: Based on the center coordinates of the set of potential hotspot areas, extract the corresponding subset of environmental parameters; The environmental parameter subset is feature-encoded to generate an environmental feature vector; The environmental feature vector is weighted and fused with the temperature anomaly probability of the potential hotspot area set; The contribution weights of environmental factors to temperature anomalies are calculated using an attention mechanism to generate a weighted fusion result. The weighted fusion result is correlated and mapped with the geographical coordinates of the potential hotspot area set to generate a comprehensive risk feature vector.

6. The temperature early warning method based on power transmission lines as described in claim 5, characterized in that, The step of using the early warning decision engine to perform hierarchical processing on the comprehensive risk feature vector includes: Analyze the temperature anomaly probability and environmental contribution weight in the comprehensive risk feature vector; Risk levels are classified according to preset risk thresholds, including low risk, medium risk, and high risk levels; Generate an early warning instruction template corresponding to each risk level, wherein the early warning instruction template includes a response action and a parameter range; The warning instruction template is dynamically adjusted based on historical warning records to generate an adaptive parameter set; Output the final temperature warning command, which integrates the risk level and adaptive parameter set.

7. The temperature early warning method based on power transmission lines as described in claim 6, characterized in that, The method further includes: During the execution of early warning commands, the temperature changes of power transmission lines are monitored in real time, and a temperature feedback data stream is generated. Compare the temperature feedback data stream with the expected results of the early warning command, and calculate the deviation index; When the deviation index exceeds the tolerance range, the model recalibration process is triggered; The parameters of the temperature anomaly detection model are updated based on the deviation index to generate an optimized model version. The optimized model version is then reapplied to temperature change feature analysis to iteratively improve the accuracy of early warning.

8. The temperature early warning method based on power transmission lines as described in claim 7, characterized in that, The triggered model recalibration process includes: The temperature feedback data stream is segmented into time windows to generate multiple monitoring time period segments; Extract temperature anomalies and their environmental context within each monitoring time segment; Calculate the similarity score between the temperature anomaly event and historical patterns to generate pattern matching results; The classification boundary of the temperature anomaly detection model is adjusted based on the pattern matching results. The model weights are updated using the gradient descent algorithm to generate a recalibrated temperature anomaly detection model.

9. The temperature early warning method based on power transmission lines as described in claim 8, characterized in that, The method further includes: Simulate environmental parameter disturbance scenarios and inject random noise into the multi-source temperature monitoring information; Observe the output stability of the early warning decision engine under disturbances and generate stability assessment data; When the stability assessment data is lower than the preset benchmark, the reinforcement learning mode of the temperature anomaly detection model is activated; The model is trained by using the difference in temperature data before and after the disturbance, which improves its robustness to abnormal patterns. Regularly verify model performance to ensure the reliability of early warning commands.

10. A temperature early warning system based on power transmission lines, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the temperature warning method based on a power transmission line as described in any one of claims 1 to 9.

Images