A power transmission line icing risk dynamic early warning method, device, equipment and medium

Abstract

The application discloses a power transmission line icing risk dynamic early warning method, device, equipment and medium, and relates to the technical field of power equipment detection. The method comprises the following steps: acquiring historical meteorological measured data of a line area, historical meteorological prediction data with a release time and an effective time mark, and original meteorological forecast data of a future period; a plurality of differential meteorological prediction sub-models generate candidate prediction results respectively, dynamic combination weights are determined according to meteorological variables and prediction lead times, and future hourly meteorological fusion prediction is formed. The fusion prediction result actually generated and archived at the historical release time is called, and a fusion prediction historical residual state feature is constructed with corresponding measured data. The state feature and the future hourly meteorological fusion prediction are paired according to the same meteorological variable and the prediction lead time, input into a time sequence risk identification model, and the future icing risk grade is output hour by hour. The application can accurately predict the icing risk of the power transmission line.

Description

Technical Field

[0001] This application relates to the field of power equipment testing technology, and in particular to a method, device, equipment and medium for dynamic early warning of icing risk on transmission lines. Background Technology

[0002] Transmission line icing is a hazardous phenomenon caused by a combination of meteorological conditions, including low temperature, high humidity, freezing rain, rime ice, and cold winds. The continuous accumulation of icing can lead to accidents such as conductor galloping, abnormal sag, insulator flashover, hardware damage, tower collapse, and line breakage, posing a serious threat to the safe and stable operation of the power grid. Therefore, developing a dynamic early warning method capable of identifying icing risks in advance has significant engineering value.

[0003] Existing icing early warning technologies mainly include methods based on empirical thresholds or physical mechanism models, methods based on on-site sensor monitoring, and methods based on numerical weather prediction or machine learning. Physical mechanism models are sensitive to local micro-topography and micro-meteorological disturbances; sensor monitoring schemes have high requirements for deployment density, communication stability, and on-site maintenance; machine learning schemes that directly use meteorological forecast values ​​are prone to passing meteorological forecast errors to the icing risk assessment stage. Summary of the Invention

[0004] The main objective of this application is to propose a method, device, equipment, and medium for dynamic early warning of icing risk of transmission lines, so as to accurately predict the icing risk of transmission lines.

[0005] To achieve the above objectives, one aspect of this application proposes a dynamic early warning method for the risk of icing on transmission lines, the method comprising the following steps: Historical meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period are acquired as meteorological data, and the meteorological data are preprocessed. Key meteorological features related to the icing status of transmission lines were screened from the preprocessed meteorological data. The key meteorological features are input into multiple differentiated meteorological prediction sub-models to obtain the sub-model prediction results of each meteorological variable under each prediction lead time. At the current moment, using only the obtained historical measured data and historical forecast data, for the two dimensions of the meteorological variable and the forecast lead, at least according to the historical forecast performance of each meteorological forecast sub-model and the complementarity between models, the dynamic combination weights are determined, and the forecast results of the sub-models are fused to obtain the future hourly meteorological fusion forecast sequence. The historical fusion prediction results are invoked; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the time series characteristics of the historical residual of fusion prediction are constructed based on the pairing results; The future hourly feature sequence is constructed by coordinating the fusion prediction historical residual time series characteristics with the future hourly meteorological fusion prediction sequence. The future hourly feature sequence is input into the time-series risk identification model, and the icing risk level of the target transmission line in the future preset time period is output hourly.

[0006] In some embodiments, the step of screening key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data includes the following steps: An improved Relief nearest neighbor feature selection algorithm based on time-series decay weights and category ratio correction is adopted to select key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The improved Relief nearest neighbor feature selection algorithm introduces time-series decay weights, category ratio correction weights, and feature dimension normalization difference in the calculation of correct and incorrect nearest neighbors to improve the contribution of samples in the critical icing period to the feature weights.

[0007] In some embodiments, the step of inputting the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecast results for each meteorological variable under each forecast lead time includes the following steps: Each of the aforementioned meteorological prediction sub-models is different in at least one of the following: model structure, input feature combination, parameter optimization method, or training mechanism; The key meteorological features are input into each of the meteorological prediction sub-models, and each of the meteorological prediction sub-models outputs the prediction results of the corresponding meteorological variables under each prediction lead.

[0008] In some embodiments, the step of determining dynamic combination weights for the meteorological variables and the forecast lead time, using only the obtained historical measured data and historical forecast data at the current moment, based at least on the historical forecast performance and inter-model complementarity of each meteorological forecast sub-model, includes the following steps: Determine the historical average absolute prediction error of each of the meteorological forecast sub-models for the meteorological variables and the forecast lead time; The historical accuracy weights are determined by normalizing the inverse of the historical average absolute prediction error. The model complementarity weights are obtained based on the correlation of the residuals of two different meteorological forecast sub-models under the same meteorological variables and the same forecast lead time. The process of fusing the prediction results of the sub-models to obtain the future hourly meteorological fusion prediction sequence includes the following steps: The weighted sum of the historical accuracy weights and the model complementarity weights is used to obtain the intermediate amount of the first combined weights; The prediction results of the sub-model are fused based on the intermediate weight of the first combination to obtain the future hourly meteorological fusion prediction sequence. The historical accuracy weight and the model complementarity weight are recalculated using historical archived data that has been verified up to the current moment when the preset rolling update cycle arrives.

[0009] In some embodiments, the method further includes the following steps: The information contribution weight is obtained based on the information entropy of the current future prediction sequence within the adjacent prediction lead amount range of the prediction lead amount; The second combined weight intermediate amount is obtained by weighting and summing the contribution weights of each information and their corresponding coefficients, and the sub-model prediction results are fused based on the second combined weight intermediate amount to update the future hourly meteorological fusion prediction sequence. When a new future prediction sequence is obtained, the information contribution weight is updated, and the dynamic combination weight is renormalized according to the preset rolling update cycle or when the information contribution weight is updated.

[0010] In some embodiments, the retrieval of historical fusion prediction results includes the following steps: The historical fusion prediction results are generated and archived based on data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the time-series features of the fusion prediction historical residuals are constructed based on the pairing results. Determine the sub-model residuals used to calculate the complementarity weights of the model; Historical fusion prediction values ​​are generated based on historical release times and data obtained prior to those times. The fusion prediction historical residual is determined based on the historical fusion prediction value and the sub-model residual; The fusion prediction historical residual temporal characteristics are constructed based on the effective time sequence of the fusion prediction historical residuals.

[0011] In some embodiments, inputting the future hourly feature sequence into the time-series risk identification model and outputting the icing risk level of the target transmission line within a preset future time period includes the following steps: The differentiated meteorological forecasting sub-model is defined as being represented by a general temporal risk identification mapping, and includes a temporal representation generation layer and a parameter-sharing classification layer. The temporal representation generation layer includes at least one of a recurrent neural network, a gated recurrent network, a temporal convolutional network, an attention-based temporal network, or a state-space sequence model, and generates a corresponding temporal representation for each input time in the future hourly meteorological fusion forecast sequence. The parameter-sharing classification layer outputs the icing risk probability corresponding to the forecast lead time hourly based on each temporal representation. The general temporal risk identification mapping serves as the temporal risk identification model. Risk classification thresholds are adaptively set based on the design ice thickness of the target transmission line, and icing risk is divided into corresponding icing risk levels based on the icing risk probability.

[0012] To achieve the above objectives, another aspect of this application proposes a dynamic early warning device for the risk of icing on transmission lines, the device comprising: The meteorological data acquisition unit is used to acquire historical measured meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period as meteorological data, and to preprocess the meteorological data. The meteorological feature screening unit is used to screen key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The meteorological forecasting unit is used to input the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecasting results of each meteorological variable under each forecast lead time. The meteorological fusion unit is used to fuse the prediction results of the sub-models at the current moment using only the obtained historical measured data and historical prediction data, and to determine the dynamic combination weights based on the historical prediction performance and inter-model complementarity of each meteorological prediction sub-model, for the two dimensions of the meteorological variable and the prediction lead, so as to obtain the future hourly meteorological fusion prediction sequence. The historical residual fusion unit is used to call historical fusion prediction results; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the fusion prediction historical residual time series features are constructed based on the pairing results; The future feature construction unit is used to collaboratively construct a future hourly feature sequence based on the fusion prediction historical residual time series features and the future hourly meteorological fusion prediction sequence. The icing risk identification unit is used to input the future hourly feature sequence into the time-series risk identification model and output the icing risk level of the target transmission line in the future preset time period on an hourly basis.

[0013] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0014] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0015] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0016] The embodiments of this application include at least the following beneficial effects: This application provides a method, device, equipment, and medium for dynamic early warning of transmission line icing risk. The solution acquires historical measured meteorological data of the target transmission line area, historical meteorological forecast data with release and effective time stamps, and future meteorological forecast data for a preset future period as meteorological data, and preprocesses the meteorological data. Key meteorological features related to the icing state of the transmission line are selected from the preprocessed meteorological data. These key meteorological features are then input into multiple differentiated meteorological forecast sub-models to obtain the sub-model prediction results for each meteorological variable under each forecast lead time. At the current moment, using only the acquired historical measured data and historical forecast data, at least based on the two dimensions of meteorological variables and forecast lead time, the system predicts the icing risk of transmission lines. The historical prediction performance and inter-model complementarity of each meteorological forecast sub-model are used to determine dynamic combination weights. The sub-model prediction results are then fused to obtain a future hourly meteorological fusion prediction sequence. Historical fusion prediction results are retrieved; these results are generated and archived based on available data at the time of historical release. The historical fusion prediction results are paired with historical measured data with the same valid time, and the time-series characteristics of the fusion prediction historical residuals are constructed based on the pairing results. The future hourly feature sequence is then constructed collaboratively with the future hourly meteorological fusion prediction sequence based on the time-series characteristics of the fusion prediction historical residuals. The future hourly feature sequence is then input into a time-series risk identification model, which outputs the icing risk level of the target transmission line within a preset future time period hourly. This application uses different meteorological forecast sub-models to predict the sub-model prediction results of each meteorological variable under each prediction lead, and then fuses the sub-model prediction results to obtain a future hourly meteorological fusion prediction sequence. Based on this future hourly meteorological fusion prediction sequence and the time-series characteristics of the fusion prediction historical residuals, an accurate future hourly feature sequence can be obtained. Therefore, the icing risk of transmission lines can be accurately predicted based on the future hourly feature sequence. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a dynamic early warning method for icing risk on transmission lines provided in this application embodiment; Figure 2 This is a flowchart illustrating the dynamic early warning process for transmission line icing risk provided in an embodiment of this application. Figure 3 A comparison chart of prediction errors of different models provided in the embodiments of this application; Figure 4 A schematic diagram comparing the accuracy of risk identification under different lead times provided in the embodiments of this application; Figure 5 A graph showing the variation of combined weights with forecast lead time provided in the embodiments of this application; Figure 6 A schematic diagram of the structure of a dynamic early warning device for icing risk of transmission lines provided in this application embodiment; Figure 7 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0020] Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows: In data-driven solutions, traditional relief feature selection mainly relies on the difference between nearest neighbors to update feature weights. If the temporal distance between samples and the proportion of risk categories are not considered, the contribution of the critical period of icing and a few risk categories can be weakened. The basic BP network is sensitive to the initial weights and thresholds. When using random search with fixed step size or fixed inertia parameters, there may be oscillations in the early search or stagnation in the later convergence. If the smoothing factor of GRNN depends on a single fixed search strategy, it may also be difficult to take into account both global search and local convergence.

[0022] The existing technology has at least four interrelated problems: First, single meteorological forecasting models have large errors in complex mountainous areas and critical icing conditions; second, different forecasting models have different predictive capabilities for meteorological variables such as temperature, wind speed, and humidity, and their relative performance varies with the forecast lead time, making it difficult to maintain stability with fixed weights or by setting only one set of weights according to the model; third, existing methods usually ignore the local micro-meteorological model mismatch information reflected in historical forecast errors and do not perform dual-time-scale co-modeling of the historical forecast residuals that have occurred and can be verified with the future meteorological forecast sequence; fourth, time series models that only output single sequence-level classification results cannot directly form hourly risk sequences that correspond one-to-one with the next H forecast times.

[0023] Existing multi-model fusion schemes typically employ global fixed weights or single objective weighting methods, making it difficult to simultaneously adapt to differences in meteorological variables and changes in forecast lead time. Some offline evaluation schemes based on residuals directly use measured values ​​at the effective forecast time to construct the current input. If these are directly applied to online early warning without historical archiving constraints, they can easily introduce information that is not yet available at the current time.

[0024] Furthermore, if a future H-hour warning is issued at the current time t, and the difference between the predicted and measured values ​​is directly used as the model input, the online warning cannot be implemented because the measured values ​​do not yet exist, and time information leakage will occur in offline experiments. Therefore, a closed-loop technical solution is needed that uses only the obtained real-time conditions to construct historical residual states at the current time, while using future weather forecasts to characterize future conditions.

[0025] This application provides a method, apparatus, device, and medium for dynamic early warning of transmission line icing risk, relating to the field of power equipment detection technology. The method, apparatus, device, and medium provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the server can also be a node server in a blockchain network; the software can be an application implementing a method for dynamic early warning of transmission line icing risk, but is not limited to the above forms.

[0026] This application can be applied to numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0027] Reference Figure 1 This application provides a dynamic early warning method for the risk of icing on transmission lines. This method may include, but is not limited to, steps S100 to S160, as follows: S100: Acquire historical meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period as meteorological data, and preprocess the meteorological data; S110: Screen key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data; S120: Input the key meteorological features into multiple differentiated meteorological prediction sub-models to obtain the sub-model prediction results of each meteorological variable under each prediction lead time; S130: At the current moment, using only the obtained historical measured data and historical forecast data, for the two dimensions of the meteorological variable and the forecast lead, at least according to the historical forecast performance and inter-model complementarity of each of the meteorological forecast sub-models, the dynamic combination weights are determined, and the forecast results of the sub-models are fused to obtain the future hourly meteorological fusion forecast sequence. S140: Call the historical fusion prediction results; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of the historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the fusion prediction historical residual time series features are constructed based on the pairing results; S150: Construct a future hourly feature sequence based on the fusion prediction historical residual time series characteristics and the future hourly meteorological fusion prediction sequence; S160: Input the future hourly feature sequence into the time-series risk identification model, and output the icing risk level of the target transmission line within a preset future time period on an hourly basis. The future input time corresponds one-to-one with the risk output time, thus obtaining the icing risk level of the target transmission line hourly.

[0028] Optionally, the step of screening key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data includes the following steps: An improved Relief nearest neighbor feature selection algorithm based on time-series decay weights and category ratio correction is adopted to select key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The improved Relief nearest neighbor feature selection algorithm introduces time-series decay weights, category ratio correction weights, and feature dimension normalization difference in the calculation of correct and incorrect nearest neighbors to improve the contribution of samples in the critical icing period to the feature weights.

[0029] Optionally, the step of inputting the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecasting results of each meteorological variable under each forecast lead time includes the following steps: Each of the aforementioned meteorological prediction sub-models is different in at least one of the following: model structure, input feature combination, parameter optimization method, or training mechanism; The key meteorological features are input into each of the meteorological prediction sub-models, and each of the meteorological prediction sub-models outputs the prediction results of the corresponding meteorological variables under each prediction lead.

[0030] Specifically, multiple meteorological forecasting sub-models are numbered j=1,…,M, where M≥2; each meteorological forecasting sub-model outputs the forecast results of the corresponding meteorological variables under each forecast lead time, and the multiple meteorological forecasting sub-models are different in at least one of the following: model structure, input feature combination, parameter optimization method or training mechanism.

[0031] Optionally, the step of determining dynamic combination weights at the current moment using only obtained historical measured data and historical forecast data, for the two dimensions of the meteorological variable and the forecast lead time, at least based on the historical forecast performance of each of the meteorological forecast sub-models and the complementarity between models, includes the following steps: Determine the historical average absolute prediction error of each of the meteorological forecast sub-models for the meteorological variables and the forecast lead time; The historical accuracy weights are determined by normalizing the inverse of the historical average absolute prediction error. The model complementarity weights are obtained based on the correlation of the residuals of two different meteorological forecast sub-models under the same meteorological variables and the same forecast lead time. The process of fusing the prediction results of the sub-models to obtain the future hourly meteorological fusion prediction sequence includes the following steps: The weighted sum of the historical accuracy weights and the model complementarity weights is used to obtain the intermediate amount of the first combined weights; The prediction results of the sub-model are fused based on the intermediate weight of the first combination to obtain the future hourly meteorological fusion prediction sequence. The historical accuracy weight and the model complementarity weight are recalculated using historical archived data that has been verified up to the current moment when the preset rolling update cycle arrives.

[0032] Specifically, the historical average absolute prediction error of the j-th meteorological forecast sub-model for meteorological variable v and prediction lead time s is: The historical accuracy weights are obtained by normalizing the inverse of the historical mean absolute prediction error. The model complementarity weights are obtained based on the correlation of the residuals of the j-th and k-th sub-models under the same meteorological variables and the same forecast lead. ;make , , where α and β are both greater than 0; the historical accuracy weight and model complementarity weight are recalculated using historical archived data that has been validated up to the current time when the preset rolling update cycle arrives.

[0033] Optionally, the method further includes the following steps: The information contribution weight is obtained based on the information entropy of the current future prediction sequence within the adjacent prediction lead amount range of the prediction lead amount; The second combined weight intermediate amount is obtained by weighting and summing the contribution weights of each information and their corresponding coefficients, and the sub-model prediction results are fused based on the second combined weight intermediate amount to update the future hourly meteorological fusion prediction sequence. When a new future prediction sequence is obtained, the information contribution weight is updated, and the dynamic combination weight is renormalized according to the preset rolling update cycle or when the information contribution weight is updated.

[0034] Specifically, the information contribution weight is obtained based on the information entropy of the current future prediction sequence within the adjacent prediction lead time range of the prediction lead time s. And adjust the intermediate weight of the combination to ,in α, β, and γ are all greater than 0, and the information contribution weight is not used alone as an evaluation of prediction reliability. The information contribution weight is updated when a new future prediction sequence is obtained, and the dynamic combination weight is renormalized according to a preset rolling update cycle or when the information contribution weight is updated.

[0035] Optionally, the step of calling historical fusion prediction results; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the time series features of the historical residuals of the fusion prediction are constructed based on the pairing results, including the following steps: Determine the sub-model residuals used to calculate the complementarity weights of the model; Historical fusion prediction values ​​are generated based on historical release times and data obtained prior to those times. The fusion prediction historical residual is determined based on the historical fusion prediction value and the sub-model residual; The fusion prediction historical residual temporal characteristics are constructed based on the effective time sequence of the fusion prediction historical residuals.

[0036] Specifically, the sub-model residual is defined as Used to calculate model complementarity weights; historical fusion prediction values ,in Based solely on historical release times τ Based on the available data for s and prior periods, historical fusion predictions are in τ. Data generated and archived at time s, without recalculating using data obtained after the effective time τ; the fused prediction historical residual is defined as... .

[0037] The fused predictions of historical residuals form a historical residual sequence in chronological order. Extract the mean, peak value, peak-to-peak value, sum of squares, and root mean square magnitude from the historical residual sequence to form... .

[0038] Optionally, the step of inputting the future hourly feature sequence into the time-series risk identification model and outputting the icing risk level of the target transmission line within a preset future time period includes the following steps: The differentiated meteorological forecasting sub-model is defined as being represented by a general temporal risk identification mapping, and includes a temporal representation generation layer and a parameter-sharing classification layer. The temporal representation generation layer includes at least one of a recurrent neural network, a gated recurrent network, a temporal convolutional network, an attention-based temporal network, or a state-space sequence model, and generates a corresponding temporal representation for each input time in the future hourly meteorological fusion forecast sequence. The parameter-sharing classification layer outputs the icing risk probability corresponding to the forecast lead time hourly based on each temporal representation. The general temporal risk identification mapping serves as the temporal risk identification model. Risk classification thresholds are adaptively set based on the design ice thickness of the target transmission line, and icing risk is divided into corresponding icing risk levels based on the icing risk probability.

[0039] Specifically, the risk of icing is divided into four levels: Level 1, Level 2, Level 3, and Level 4.

[0040] The following sections will provide a detailed description and explanation of some optional embodiments of this application, using specific application examples.

[0041] The purpose of this application is to provide a dynamic early warning method for transmission line icing risk based on the fusion prediction of historical meteorological forecast residuals and future meteorological forecasts, in order to solve the problems of future residuals not being available in real time, different meteorological variables and different forecast lead times sharing the same model weights, confusion between the meanings of the two types of residuals, and the lack of hourly corresponding output capability for the next H risk times.

[0042] Figure 2 This application presents a flowchart for the dynamic early warning system for transmission line icing risk. The embodiment employs a dual-timescale technology chain: on the historical timescale, historical forecasts with corresponding measured values ​​are paired according to meteorological variables and forecast lead times, constructing sub-model residuals and fused forecast historical residuals respectively; on the future timescale, model combination weights W are determined for each meteorological variable v and each forecast lead time s. j,v,s (t) is used to obtain the hourly meteorological fusion forecast for the future; the historical fusion residual state under the same meteorological variables and the same forecast lead time is input together with the future fusion meteorological conditions into the time series risk identification model, and the future icing risk is output hourly.

[0043] Let the current time be t, there are M meteorological forecast sub-models, V key meteorological variables, and a future forecast time domain length of H. The forecast lead time s∈{1,…,H}. Let j=1,…,M represent the sub-model numbers. Each sub-model uses a unified symbol to represent the forecast result, and the mathematical expression is not limited by the specific model type. j,v,t+s|t x represents the predicted value of the j-th sub-model at time t for the meteorological variable v, with an effective time of t+s; v,τ This represents the measured value at the valid time τ. For a historical time τ that has been validated, the sub-model residual and the fusion prediction historical residual are defined as follows: ; ; ; Where, r sub Used only for calculating complementarity between sub-models, r fus This describes the historical adaptation status of fused meteorological forecasts to the local micro-meteorological environment along the railway line. Historical fused forecast value D v,τ|τ s At the moment of historical release τ s is based on the data available at the time and the dynamically combined weights W determined at that time. j,v,s (τ s) Real-time generated and archived forecast results; archived content includes at least the forecast release time, forecast validity period, meteorological variables, forecast lead time, forecast values ​​of each sub-model, combined weights, and fused forecast values. Data obtained after the validity period τ is not recalculated. Let t a ≤t represents the latest valid time when measured values ​​can be obtained at the current time; for each meteorological variable v and forecast lead s, select the most recent historical fusion residuals up to the latest valid time when measured values ​​can be obtained in the order of valid time, form a historical fusion residual sequence, and extract the time series state features: ; ; For each future forecast time t+s, the fused forecasts of V meteorological variables, the V sets of historical fused residual states matching the forecast lead, and the normalized lead s / H are combined into an hourly input z. t+s Thus, the future input sequence Z is constructed. t : ; ; ; Where G represents a general temporal risk identification mapping, used to maintain a one-to-one correspondence between future input times and risk output times; G is not limited by a specific temporal network structure. Online construction of Z t Only historical measured data and future weather forecast data obtained at the current time are used, and weather measured values ​​x that have not yet been obtained at future forecast times are not used. v,t+1 ,…,x v,t+H The measured values ​​will only be used to update the historical fusion residual sequence and generate offline training labels for subsequent warning times after the corresponding valid time arrives and is verified.

[0044] Meteorological variables—adaptive dynamic weighting of forecast lead time: The historical performance evaluation period includes the most recent N valid time intervals that have been verified, and its endpoint is the latest valid time interval t at which the measured value can be obtained at the current time. a The weights for the historical average absolute prediction error and historical accuracy of the j-th sub-model for the meteorological variable v and the prediction lead s are as follows: ; ; Therefore, when predicting different variables such as temperature, wind speed, and humidity, each model is weighted according to the historical performance of its corresponding variable; when predicting different time periods such as t+1 and t+H, each model is weighted according to the historical performance of the same prediction lead.

[0045] Model redundancy is measured by the residual correlation of sub-models under the same meteorological variables and the same forecast lead. The average correlation, independence index, and model complementarity weights of the j-th sub-model are as follows: ; ; ; The lower the correlation between the residuals of a model and its sub-models, the more complementary information it provides, and thus the higher its model complementarity weight.

[0046] The basic implementation method combines historical accuracy weights and model complementarity weights at least, and first calculates the intermediate quantity Q of the combined weights. j,v,s (t), then normalized to obtain the dynamic combined weights: ; ; In the enhanced implementation, a third type of weight can also be introduced to describe the contribution of future predicted sequence information. This weight does not evaluate prediction reliability independently. Taking H≥2, a set S of adjacent prediction lead times is selected for each prediction lead time s. s ={q|1≤q≤H and|q s|≤b}, where the radius of the adjacent forecast lead range b≥1, so that |S s |≥2; First, perform non-negative normalization on the prediction sequence of the j-th sub-model and meteorological variable v within the adjacent prediction lead range, then calculate the local information entropy, difference coefficient, and information contribution weight: ; ; ; ; Information contribution weights are used only in combination with historical accuracy weights and model complementarity weights. The combination coefficients are determined using the training and validation sets, and α ≥ γ can be set to avoid highly volatile but low-accuracy prediction sequences receiving excessive weights simply because of large information differences. For prediction values ​​exceeding the historical reasonable fluctuation range, amplitude limiting or anomaly detection is first performed. When the local fluctuation amplitude of the current prediction sequence exceeds the reasonable upper limit determined by the training or validation set, its non-negative normalized out-of-bounds amplitude relative to the reasonable upper limit is used as Δ, and the information contribution weights are corrected using an anomaly attenuation factor. The anomaly attenuation factor and the corrected information contribution weights are as follows: ; ; Where Δ is the non-negative normalized out-of-bounds amplitude of the local fluctuation relative to the reasonable upper limit, η>0 is the attenuation intensity parameter, and λ is the abnormal attenuation factor; when the local fluctuation amplitude does not exceed the reasonable upper limit, Δ is 0 and λ is 1. The corrected information contribution weights are re-normalized and then used in the calculation of the combined weights, so that the information contribution ratio of the abnormally high volatility sequence decreases monotonically as the out-of-bounds amplitude increases. In this enhanced implementation, the intermediate amount of the combined weights is adjusted as follows: ; ; ; In the above calculations, the historical accuracy weight and model complementarity weight use only historical data that has been validated up to the current moment, and the information contribution weight uses only the currently available future prediction sequences. The entire fusion process does not depend on future measured values.

[0047] The technical implications of historical residual status: Under complex terrain and icing-critical meteorological conditions, a systematic mismatch can easily occur between local micrometeorological conditions and conventional meteorological forecasting models. This mismatch manifests over time as a continuous shift, increased fluctuation amplitude, or abrupt changes in the fused forecast historical residuals. fus v,s(t) does not directly characterize the icing thickness, but is used to describe the historical fit of the fused meteorological forecast to the local micro-meteorological environment of the current line under meteorological variable v and forecast lead s, and is used together with the future meteorological forecast features with the same variable and forecast lead for icing risk identification.

[0048] Icing risk classification: Using equivalent icing thickness h as the historical training label and risk classification basis, thresholds T1, T2, and T3 are set, and the risk level is defined as follows: ; For lines in different areas, the ice thickness H is determined according to the design requirements. d The optimal conversion relationship for adaptively setting the tiered thresholds is as follows: ; This classification avoids explicitly stating that a small icing thickness is "no icing" and allows the risk level to be adapted to lines with different designed icing thicknesses.

[0049] Symbol and parameter descriptions: The symbols, value ranges, and technical meanings used in the embodiments of this application are uniformly as follows; unless otherwise specified, the specific values ​​in the embodiments are used to describe the executable configuration.

[0050] Table 1. Global Variables and Symbols Correspondence

[0051] The beneficial effects of the embodiments of this application are as follows: The embodiments of this application have data dependencies with sequential constraints in each technical step: the two-dimensional dynamic combination weight W of meteorological variable and forecast lead time is first generated and archived according to the release time and effective time of the fusion prediction D; after the corresponding measured value is reached, the archived D can form the historical residual of the fusion prediction and its state characteristics F; F is then paired item by item with the current future fusion prediction D with the same meteorological variable and the same forecast lead time to form the hourly input of the general time series risk identification mapping G and output the corresponding risk level. The result of any preceding step will change the subsequent input, so this scheme forms a closed-loop data chain without leakage of future information, rather than a simple parallel of multi-model fusion, residual analysis and time series classification.

[0052] 1) Dual time scale architecture: Online early warning does not call future measured values ​​and prohibits the use of post-event data to calculate historical fusion predictions, so that residual status can be directly used for real-time rolling early warning.

[0053] 2) Two-dimensional dynamic fusion of meteorological variables and forecast lead: The combined weights are refined to two dimensions: meteorological variables and forecast lead. This allows different meteorological variables and different forecast lead to select appropriate model combinations, which helps to reduce the systematic forecast errors caused by fixed and uniform weights.

[0054] 3) Two types of residuals are used separately: strictly distinguish between the sub-model residuals used to calculate model complementarity and the fusion prediction historical residuals used for risk identification, while using error information to optimize fusion and characterize local micro-meteorological mismatch.

[0055] 4) Multi-input multi-output time-series risk identification model: This model ensures that multiple future meteorological input times correspond one-to-one with multiple icing risk output times, meeting the hourly output requirements of the power grid rolling early warning service.

[0056] 5) Center-edge collaborative deployment: Supports training of the central server and inference of the substation edge terminal, and maintains the most recent effective weight and executes local degradation warning when communication is interrupted, thereby improving the continuous operation capability of the project.

[0057] 6) Adaptive risk grading: Based on the ice thickness of the target line design, four levels of risk thresholds are adaptively set to avoid insufficient adaptation of fixed absolute thresholds to lines with different ice thicknesses.

[0058] In the above-described solutions of the embodiments of this application, some optional drawings are included: Figure 3 This is a comparison chart of the prediction errors of different models. Figure 4 This diagram illustrates the comparison of risk identification accuracy under different forecast lead times. Figure 5 This is a graph showing how the combined weights change with the forecast lead time.

[0059] The embodiments of this application will be further described below with reference to specific examples.

[0060] Example 1: A dynamic early warning method for icing risk of transmission lines based on multi-model fusion and predictive residual time series state analysis.

[0061] S1, Data Acquisition and Preprocessing. Acquire hourly historical meteorological data for the target route area, historical meteorological forecast data with release and effective time stamps, future H-hour meteorological forecast data, and corresponding historical icing monitoring tags. Meteorological elements include at least ambient temperature, daily minimum temperature, relative humidity, saturated humidity, absolute humidity, and wind speed. The Raida criterion is used to identify candidate outliers, and the change rate of adjacent time intervals, historical quantiles for the same period, and sensor status are combined to distinguish between true extreme values ​​and mismeasured values. A missing mask is set for mismeasured or missing time intervals. Interpolation of adjacent time intervals or historical averages under the same conditions are only used to maintain the continuity of non-residual inputs and are not used as true measured values ​​in historical accuracy weights, model complementarity weights, or calculations of fusion prediction historical residuals. The normalization formula is: ; Where, x min and x maxIt is determined only by the training set and remains fixed in the validation set, test set, and online phase to avoid data leakage.

[0062] S2, Improved Relief Feature Selection. Preprocessed data is input into an improved Relief nearest neighbor feature selection algorithm based on temporal decay weights and class proportion correction. The normalized dissimilarity of feature A, the temporal decay weight of the l-th nearest neighbor, and the feature weight increment are as follows: ; ; ; Where R is the sample size, h l To guess the nearest neighbor, s l To determine the incorrect nearest neighbor, k represents the number of nearest neighbors, and τ0 represents the temporal decay scale. Compared to the Relief update method that does not consider temporal relationships and category proportions, the temporal decay weight increases the contribution of icing evolution samples that are close to the sampling time, while category proportion correction reduces the dominant role of most risk categories in feature weight updates. Based on the training set calculation results, at least two of the following categories—wind speed, ambient temperature, daily minimum temperature, and relative humidity—are selected to constitute the key meteorological feature set; the specific selection results are based on the training data of the actual deployment area.

[0063] S3, Construction of Differentiated Weather Forecast Sub-models. The output of the j-th weather forecast sub-model is uniformly denoted as the j-th sub-model prediction value, j=1,…,M; each model differs in at least one of the following: model structure, input feature combination, parameter optimization method, or training mechanism. In this embodiment, M=3, and the following sub-models are adopted: a backpropagation neural network (BP) sub-model optimized by a random walk with decreasing inertial weights, a generalized regression neural network (GRNN) sub-model optimized by a modified swarm search algorithm for smoothing factors, and an extreme learning machine (ELM) sub-model. The BP sub-model uses a random walk with decreasing inertial weights to optimize the initial weights and thresholds; the GRNN sub-model uses a modified swarm search algorithm to optimize the smoothing factor SPREAD; the ELM sub-model randomly generates the connection weights and biases from the input layer to the hidden layer, and solves the output weights using the least squares method. BP, GRNN, and ELM are only the model selections for this embodiment; other models that can output prediction results under the same meteorological variables and prediction lead can also be used as the j-th weather forecast sub-model.

[0064] ; ; ; ; In the BP sub-model optimization formula, Bi t Let B represent the i-th candidate initial weight and threshold vector in the t-th iteration. best t B represents the current best candidate. rand t This represents random candidates, where r1 and r2 are independent random numbers within [0,1]. `clip` restricts the candidates to preset upper and lower bounds. ω max and ω min Here, X represents the initial and final inertia weights, respectively, and T represents the maximum number of iterations. In the GRNN sub-model optimization formula, X... j t Let X represent the j-th candidate smoothing factor. best t Lb represents the current optimal smoothing factor. j and Ub j To search for the lower and upper bounds, s j For taking values Search direction of 1 or 1, VWEP t VTDR represents the global candidate proximity probability that increases with iteration. t p is the decay exponent, representing the search distance rate that decreases with iteration.

[0065] ; Where ξ is the perturbation quantity following a standard normal distribution, η0 is the initial perturbation amplitude, q is the number of consecutive iterations in which the objective function on the validation set fails to improve, and q0 is the stagnation threshold. Candidate solutions are uniformly initialized within the upper and lower bounds of the parameters, and the mean absolute error of the validation set is used as the objective function. The optimization terminates when the maximum number of iterations T is reached, or when the relative change in the objective function is less than the preset threshold for q0 consecutive times.

[0066] Compared to directly and randomly initializing BP parameters or using a fixed inertia parameter for random search, decreasing the inertia weights allows the optimization process to gradually transition from a strong global exploration in the early stages to local convergence in the later stages. Compared to optimizing the GRNN smoothing factor only by a fixed search distance, the global candidate proximity probability, search distance rate, and stagnation perturbation that change with iterations together reduce the risk of premature convergence or continuous stagnation.

[0067] S4, Adaptive Dynamic Fusion of Meteorological Variables and Forecast Lead. For each meteorological variable v and each forecast lead s, at least according to the aforementioned historical accuracy weight w... acc j,v,s Complementary weights w of the model ind j,v,s The computation chain obtains W j,v,s (t), and further combined with the information contribution weight w info j,v,s Calculate D v,t+s|tThe combination coefficients and the radius b of the adjacent forecast lead range are determined through the training and validation sets. During the online phase, the weights of the first two categories are updated using the rolling historical records of the same variables and the same forecast lead, and the current future forecast sequence is used to update the contribution weights when the enhanced implementation is enabled.

[0068] S5, Two types of residual construction. For each valid time τ where measured values ​​have been obtained up to the current time t, the historical publication time τ is called. Sub-model predictions generated and archived in real time j,v,τ|τ s Dynamic combination weights W j,v,s (τ s) and fusion prediction value D v,τ|τ s The archived value is only composed of τ. Data available at time s and before are determined; data obtained after time τ are not used for retrospective calculations. r is calculated separately. sub j,v,τ,s and r fus v,τ,s The former is only used to update the model's complementarity weights, while the latter is used to construct historical mismatch states for risk identification. It is prohibited to construct the current input using measured values ​​from the next H hours; when there is a communication delay with the latest measured value, the latest available measured time t should be used. a ≤t is the termination time of the historical residual sequence.

[0069] S6, Constructing the Fusion Residual Time Series Status and Risk Input. For each meteorological variable v and forecast lead s, select the most recent m fused forecast historical residuals up to the latest available measured time in effective time order to form the sequence R. fus v,s (t), and extract the following five types of features: ; ; ; ; The fusion prediction of V meteorological variables at time t+s, the V sets of fusion residual state characteristics corresponding to the prediction lead s, and s / H are concatenated into z. t+s And arrange them in order of s=1,…,H to obtain Z t For specific parameter configurations, please refer to Example 2.

[0070] S7, hourly identification and release of the time-series risk identification model. (Z) tThe general temporal risk identification mapping G is input in future time sequence. G outputs the corresponding temporal representation h at each input position t+s. t+s Instead of using only the last hidden state, each temporal representation is fed into a parameter-shared fully connected layer and a Softmax layer to obtain the risk probability π for the corresponding time step. t+s,c and risk level L t+s This will form an hourly risk sequence for the next H hours.

[0071] ; ; ; Training labels are generated from icing monitoring results corresponding to each future valid moment in historical time periods, and missing labels are excluded using a mask. In this embodiment, G adopts an LSTM structure; other implementations may use GRU, temporal convolution, attention, or state-space sequence structures, as well as other structures that can generate temporal representations position-by-position and connect them to share a classification layer, none of which constitute a limitation on G. Online inference only uses features available up to the current moment, and level one to four risk warnings can be issued via human-computer interaction display devices, audible and visual alarm devices, or power dispatching systems.

[0072] Example 2: Executable parameter configuration and tensor construction.

[0073] In this embodiment, the number of models, prediction time domain, length of historical performance evaluation period, length of fused prediction historical residual sequence, combination coefficient, network size, training rounds, and other values ​​are all preferred implementation parameters used to illustrate the executable configuration. They do not constitute unique values ​​and can be adaptively adjusted through the training set and validation set according to the line monitoring frequency, design ice thickness, regional climate characteristics, edge computing power, and early warning latency requirements.

[0074] This embodiment exemplifies a directly achievable parameter configuration. Ambient temperature, daily minimum temperature, relative humidity, and wind speed are selected as the four key meteorological variables (V=4). The number of meteorological prediction sub-models is set to M=3, the future prediction time domain to H=48 hours, the length of the fused prediction historical residual sequence to m=12 hours, the length of the historical performance evaluation period to N=720 hours, the radius of the adjacent prediction lead range to b=2, and the numerical stability constant to ε=10. 8 The basic dynamic fusion uses α=0.60 and β=0.40; when information contribution enhancement is enabled, α=0.55, β=0.35, and γ=0.10 are used.

[0075] The BP sub-model optimized using a random walk with decreasing inertia weights is configured with a candidate population size of 30, a maximum number of iterations T=100, and ω... max =0.9、ωmin =0.4. The GRNN sub-model optimized using the improved group search algorithm sets the SPREAD search interval to [0.01, 2.00], the number of candidates to 30, the maximum number of iterations T=100, and VWEP. min =0.2, VWEP max =1.0, decay exponent p=6, stagnation threshold q0=10, initial disturbance amplitude η0=0.05 (Ub Lb). When the relative change in the mean absolute error of the validation set is less than 10 for 10 consecutive times. 6 The optimization process may terminate after 100 iterations.

[0076] Each forecast lead s corresponds to an hourly input containing V fused meteorological forecast values, 5 fused residual state features for each meteorological variable, and 1 normalized lead. Therefore, the hourly input dimensions are: ; ; In this embodiment, the general temporal risk identification mapping G specifically adopts a two-layer LSTM, with 64 hidden units in each layer and a dropout ratio of 0.2 between layers. The second layer outputs 64-dimensional hidden states at 48 input positions. Each hidden state enters the same fully connected classification layer with shared parameters. The classification layer maps the 64-dimensional hidden states into logical values ​​of 4 risk levels, and then forms a time-series risk probability matrix through Softmax. ; Training was performed using a batch size of 64, the Adam optimizer, and an initial learning rate of 10. 3 L2 regularization coefficient 10 4 The maximum number of training epochs is 100. Early stopping is achieved by using the cross-entropy of the validation set sequence as the criterion; training stops and the parameters for the lowest validation set loss are restored if the cross-entropy does not decrease for 10 consecutive epochs. In cases of class imbalance, inverse frequency class weights are calculated only based on the frequency of samples at each risk level in the training set. During the online phase, a 48×25 input tensor is constructed every hour, and a 48×4 risk probability matrix is ​​output. Simultaneously, the predictions of each sub-model at that moment are archived. j,v,s (t) and D v,t+s|t This is used for updating historical residuals after the corresponding measured values ​​are reached.

[0077] Example 3: Implementation plan for training, verification and control experiments.

[0078] The training dataset consists of historical meteorological data of the target transmission line or similar lines, historical meteorological forecast data retaining the release time and validity period, and corresponding icing monitoring labels. Each training record includes at least the line identifier, meteorological station or grid identifier, forecast release time, forecast validity period, forecast lead time, meteorological variables, measured values, and risk labels. Equivalent icing thickness can be obtained from at least one of online tension / weighing devices, image or fiber optic monitoring devices, or manual observation records. Meteorological forecast data can come from numerical weather prediction systems, line micro-meteorological forecast systems, or fusion forecast results output by the above systems.

[0079] To avoid time leakage, the training, validation, and test sets are divided according to the icing events and their chronological order. Adjacent time series samples generated during the same icing process cannot be included in both the training and test sets simultaneously. Normalization parameters, improved Relief feature selection results, combination coefficients, radius b of the adjacent prediction lead range, length m of the fused prediction historical residual sequence, number of network layers, number of hidden units, and regularization parameters are all determined solely from the training set or the training and validation sets.

[0080] To verify the role of historical fusion residual state, two-dimensional dynamic fusion of meteorological variables and forecast lead time, and information contribution enhancement terms in icing risk identification, each ablation group used the exact same key feature set, meteorological forecast sub-model and its output, risk identification network structure, parameter initialization rules, event-level time series partitioning of training, validation and test sets, normalization parameters, batch size, maximum training rounds, early stop rules and evaluation indicators, and only switched the variable of historical residual input or combined weight type, and set up the following control experiment.

[0081] Table 2 Input and Dynamic Fusion Comparison Design

[0082] Experimental groups A and C use identical sub-models, fixed model weights, data partitioning, training epochs, and risk identification networks. The only difference between them is whether historical fusion residual states are input. Therefore, the difference between C and A is used to evaluate the incremental effect of historical residual states. Experimental groups C and D maintain the same conditions except for fixed and dynamic weights. Therefore, the difference between D and C is used to evaluate the incremental effect of adaptive dynamic fusion of meteorological variables and forecast lead. Experimental group B is used to evaluate the independent information contribution of historical fusion residual states when future meteorological forecasts are not input.

[0083] The three sets of experiments should be compared using the same data, the same sub-model predictions, and the same risk identification network, and the weather forecast error and icing risk index should be reported separately according to the forecast lead time. Only when P+I+Info outperforms P+I on the independent test set should the information contribution weight be considered to have a positive effect; otherwise, γ should be reduced or this item should be discontinued.

[0084] Regional assessment terminology is determined based on training and testing methods: if each region is trained and tested separately, it should be called "different-region applicability testing"; if multiple regions are trained together and statistics are collected separately, it should be called "regional performance statistics"; only when training is conducted in one or more source regions and the target regions not involved in the training are they called "cross-regional generalization testing". Cross-regional generalization testing collects statistics on the macro average F1, recall rate, false alarm rate, false alarm rate, and performance changes relative to the source region test results for each target region.

[0085] Extreme scenario tests can select freezing rain, persistent rime, and low-temperature strong wind events, and include them in the test set as a whole. The event-level detection rate, the warning lead time for the first time reaching level three or above risk, the deviation of the highest risk level, and the number of false alarms during the event are statistically analyzed. The technical explanation for the low-temperature strong wind scenario is as follows: wind speed, as a key meteorological feature selected by the improved Relief algorithm, can enhance the model's ability to represent the state of low-temperature strong wind icing risk scenarios; without setting an online feature weight update mechanism, no conclusion is drawn regarding online changes to the wind speed feature weight.

[0086] Example 4: Lightweight deployment of edge computing power. When deploying low-computing-power terminals locally at substations, the training of various meteorological forecast sub-models and multi-input multi-output time-series risk identification models can be completed on the central server, while the field terminals only perform forward forecasting and W... j,v,s (t) Update and fuse the calculation of historical residual states and hourly risk level output. The multi-input multi-output time series risk identification model can use a single-layer LSTM or GRU and reduce the number of hidden units; when the early warning delay requirement is high, the length of the fused prediction historical residual sequence and the length of the future prediction time domain can be adjusted according to the validation set results.

[0087] The above scheme is tested in the same edge terminal and runtime environment using the same test data, batch size and risk classification rules; the P50 and P95 inference latency, model loading and peak memory during a single inference are recorded after warm-up, and the macro average F1, false alarm rate and false alarm rate are statistically analyzed according to the event-level test set to quantify the resource benefits and risk identification performance changes.

[0088] The lightweight configuration is compared with the standard time series model using an independent test set in terms of macro-average F1, recall rate for level 3 and 4 risks, false alarm rate, inference latency, and peak memory usage. The engineering acceptance threshold is determined based on line voltage level, risk tolerance, early warning latency requirements, and terminal computing power, combined with actual measurement results. When any key risk identification indicator or resource indicator fails to meet the engineering requirements of the corresponding line, the standard time series model is retained, or the scale of the time series representation generation layer, the length of the fused prediction historical residual sequence, and the length of the future prediction time domain are readjusted.

[0089] If communication is interrupted, the terminal retains the most recent valid model parameters and combined weights, and marks the output as degraded mode; after communication is restored, the measured values ​​will be retransmitted and paired with the corresponding historically published prediction values ​​to supplement the generation of fusion prediction historical residuals for the missing time period; this result is only used for rolling early warning after recovery and must not rewrite the previous online input or early warning results.

[0090] When on-site meteorological sensors fail, the terminal identifies missing variables based on the status code. Missing time points cannot be used to update historical accuracy weights, model complementarity weights, or fusion prediction historical residuals; instead, they are skipped using a missing mask, and the most recent valid measured time is used as the termination time of the historical residual sequence. The terminal can use still available numerical weather predictions and interpolate based on historical micrometeorological conditions determined from training data and adjacent valid time points to maintain continuity of non-residual inputs. The interpolation results cannot be used as measured values ​​in historical residual archiving or model performance evaluation, and the risk output of affected time points is marked as low confidence.

[0091] When communication interruption lasts for more than one data update cycle, the terminal retains the most recently valid dynamic combination weight and uses locally stored and still valid weather forecasts and historical fusion residual states to perform short-term risk extrapolation; after the weather forecast fails, it stops outputting deterministic long-term levels and only outputs short-term downgrade warnings based on local residual states and marks the data timeliness.

[0092] When extreme icing conditions such as freezing rain, persistent rime, or continuous unidirectional shift of residuals are detected, the length m of the fused prediction historical residual sequence is adjusted within a range of 6 to 24 hours; the specific length adjustment rules are determined by the training set and validation set, and the normal sequence length is restored after the event ends.

[0093] Reference Figure 6 This application also provides a dynamic early warning device for transmission line icing risk, which can realize the above-mentioned dynamic early warning method for transmission line icing risk. The device includes: The meteorological data acquisition unit is used to acquire historical measured meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period as meteorological data, and to preprocess the meteorological data. The meteorological feature screening unit is used to screen key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The meteorological forecasting unit is used to input the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecasting results of each meteorological variable under each forecast lead time. The meteorological fusion unit is used to fuse the prediction results of the sub-models at the current moment using only the obtained historical measured data and historical prediction data, and to determine the dynamic combination weights based on the historical prediction performance and inter-model complementarity of each meteorological prediction sub-model, for the two dimensions of the meteorological variable and the prediction lead, so as to obtain the future hourly meteorological fusion prediction sequence. The historical residual fusion unit is used to call historical fusion prediction results; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the fusion prediction historical residual time series features are constructed based on the pairing results; The future feature construction unit is used to collaboratively construct a future hourly feature sequence based on the fusion prediction historical residual time series features and the future hourly meteorological fusion prediction sequence. The icing risk identification unit is used to input the future hourly feature sequence into the time-series risk identification model and output the icing risk level of the target transmission line in the future preset time period on an hourly basis.

[0094] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0095] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method of this application. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0096] It is understood that the content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the methods of this application, and the beneficial effects achieved are the same as those achieved by the methods of this application.

[0097] Figure 7 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 101 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 102 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 102 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 102 and is called and executed by the processor 101. Input / output interface 103 is used to implement information input and output; The communication interface 104 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 105 transmits information between various components of the device (e.g., processor 101, memory 102, input / output interface 103, and communication interface 104); The processor 101, memory 102, input / output interface 103 and communication interface 104 are connected to each other within the device via bus 105.

[0098] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of this application.

[0099] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0100] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0101] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0102] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0103] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0104] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0105] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0106] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0107] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0109] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0110] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0111] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A dynamic early warning method for the risk of icing on transmission lines, characterized in that, The method includes the following steps: Historical meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period are acquired as meteorological data, and the meteorological data are preprocessed. Key meteorological features related to the icing status of transmission lines were screened from the preprocessed meteorological data. The key meteorological features are input into multiple differentiated meteorological prediction sub-models to obtain the sub-model prediction results of each meteorological variable under each prediction lead time. At the current moment, using only the obtained historical measured data and historical forecast data, for the two dimensions of the meteorological variable and the forecast lead, at least according to the historical forecast performance of each meteorological forecast sub-model and the complementarity between models, the dynamic combination weights are determined, and the forecast results of the sub-models are fused to obtain the future hourly meteorological fusion forecast sequence. The historical fusion prediction results are invoked; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the time series characteristics of the historical residual of fusion prediction are constructed based on the pairing results; The future hourly feature sequence is constructed by coordinating the fusion prediction historical residual time series characteristics with the future hourly meteorological fusion prediction sequence. The future hourly feature sequence is input into the time-series risk identification model, and the icing risk level of the target transmission line in the future preset time period is output hourly.

2. The method for dynamic early warning of icing risk on transmission lines according to claim 1, characterized in that, The process of screening key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data includes the following steps: An improved Relief nearest neighbor feature selection algorithm based on time-series decay weights and category ratio correction is adopted to select key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The improved Relief nearest neighbor feature selection algorithm introduces time-series decay weights, category ratio correction weights, and feature dimension normalization difference in the calculation of correct and incorrect nearest neighbors to improve the contribution of samples in the critical icing period to the feature weights.

3. The method for dynamic early warning of icing risk on transmission lines according to claim 1, characterized in that, The process of inputting the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecast results for each meteorological variable under each forecast lead time includes the following steps: Each of the aforementioned meteorological forecast sub-models is to be different in at least one of the following: model structure, input feature combination, parameter optimization method, or training mechanism; The key meteorological features are input into each of the meteorological prediction sub-models, and each of the meteorological prediction sub-models outputs the prediction results of the corresponding meteorological variables under each prediction lead.

4. The method for dynamic early warning of icing risk on transmission lines according to claim 1, characterized in that, The method of determining dynamic combination weights for the meteorological variables and the forecast lead time, using only the historical measured data and historical forecast data obtained at the current moment, based at least on the historical forecast performance and inter-model complementarity of each meteorological forecast sub-model, includes the following steps: Determine the historical average absolute prediction error of each of the meteorological forecast sub-models for the meteorological variables and the forecast lead time; The historical accuracy weights are determined by normalizing the inverse of the historical average absolute prediction error. The model complementarity weights are obtained based on the correlation of the residuals of two different meteorological forecast sub-models under the same meteorological variables and the same forecast lead time. The process of fusing the prediction results of the sub-models to obtain the future hourly meteorological fusion prediction sequence includes the following steps: The weighted sum of the historical accuracy weights and the model complementarity weights is used to obtain the intermediate amount of the first combined weights; The prediction results of the sub-model are fused based on the intermediate weight of the first combination to obtain the future hourly meteorological fusion prediction sequence. The historical accuracy weight and the model complementarity weight are recalculated using historical archived data that has been verified up to the current moment when the preset rolling update cycle arrives.

5. The method for dynamic early warning of icing risk on transmission lines according to claim 4, characterized in that, The method further includes the following steps: The information contribution weight is obtained based on the information entropy of the current future prediction sequence within the adjacent prediction lead amount range of the prediction lead amount; The second combined weight intermediate amount is obtained by weighting and summing the contribution weights of each information and their corresponding coefficients, and the sub-model prediction results are fused based on the second combined weight intermediate amount to update the future hourly meteorological fusion prediction sequence. When a new future prediction sequence is obtained, the information contribution weight is updated, and the dynamic combination weight is renormalized according to the preset rolling update cycle or when the information contribution weight is updated.

6. The method for dynamic early warning of icing risk on transmission lines according to claim 4, characterized in that, The process involves calling historical fusion prediction results; wherein the historical fusion prediction results are generated and archived based on data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the time series features of the historical residuals of the fusion prediction are constructed based on the pairing results, including the following steps: Determine the sub-model residuals used to calculate the complementarity weights of the model; Historical fusion prediction values ​​are generated based on historical release times and data obtained prior to those times. The fusion prediction historical residual is determined based on the historical fusion prediction value and the sub-model residual; The fusion prediction historical residual temporal characteristics are constructed based on the effective time sequence of the fusion prediction historical residuals.

7. A dynamic early warning method for icing risk of transmission lines according to any one of claims 1 to 6, characterized in that, The step of inputting the future hourly feature sequence into the time-series risk identification model and outputting the icing risk level of the target transmission line within a preset future time period includes the following steps: The differentiated meteorological forecasting sub-model is defined as being represented by a general temporal risk identification mapping, and includes a temporal representation generation layer and a parameter-sharing classification layer. The temporal representation generation layer includes at least one of a recurrent neural network, a gated recurrent network, a temporal convolutional network, an attention-based temporal network, or a state-space sequence model, and generates a corresponding temporal representation for each input time in the future hourly meteorological fusion forecast sequence. The parameter-sharing classification layer outputs the icing risk probability corresponding to the forecast lead time hourly based on each temporal representation. The general temporal risk identification mapping serves as the temporal risk identification model. Risk classification thresholds are adaptively set based on the design ice thickness of the target transmission line, and icing risks are classified into corresponding icing risk levels based on the icing risk probability.

8. A dynamic early warning device for the risk of icing on transmission lines, characterized in that, The device includes: The meteorological data acquisition unit is used to acquire historical measured meteorological data of the target transmission line area, historical meteorological forecast data with release time and effective time stamps, and future meteorological forecast data for a preset future period as meteorological data, and to preprocess the meteorological data. The meteorological feature screening unit is used to screen key meteorological features related to the icing status of transmission lines from the preprocessed meteorological data. The meteorological forecasting unit is used to input the key meteorological features into multiple differentiated meteorological forecasting sub-models to obtain the sub-model forecasting results of each meteorological variable under each forecast lead time. The meteorological fusion unit is used to fuse the prediction results of the sub-models at the current moment using only the obtained historical measured data and historical prediction data, and to determine the dynamic combination weights based on the historical prediction performance and inter-model complementarity of each meteorological prediction sub-model, for the two dimensions of the meteorological variable and the prediction lead, so as to obtain the future hourly meteorological fusion prediction sequence. The historical residual fusion unit is used to call historical fusion prediction results; wherein, the historical fusion prediction results are generated and archived based on the data available at the time of historical release; the historical fusion prediction results are paired with historical measured data with the same valid time, and the fusion prediction historical residual time series features are constructed based on the pairing results; The future feature construction unit is used to collaboratively construct a future hourly feature sequence based on the fusion prediction historical residual time series features and the future hourly meteorological fusion prediction sequence. The icing risk identification unit is used to input the future hourly feature sequence into the time-series risk identification model and output the icing risk level of the target transmission line in the future preset time period on an hourly basis.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images