Power transmission line icing prediction method based on time-space diagram topology and micrometeorological feature fusion

Abstract

The invention discloses a power transmission line icing prediction method based on time-space diagram topology and micrometeorological feature fusion, and relates to the technical field of power transmission line icing prediction. A graph model is constructed through icing thickness correlation among nodes, a spatial relationship among the nodes and correlation among meteorological factors; then, feature extraction is carried out on the graph model through a GCN model, and spatial features among all nodes are obtained; and the extracted spatial features are input into the LSTM-SNPS to perform time sequence feature extraction and complete icing thickness prediction, so that the problem of time sequence prediction deviation caused by spatial information loss can be avoided, and the accuracy of icing thickness prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of transmission line icing prediction technology, specifically to a method for predicting transmission line icing based on the fusion of spatiotemporal map topology and micro-meteorological features. Background Technology

[0002] Icing on transmission lines is a common natural disaster in power systems, and severe icing can lead to line overload accidents. Furthermore, uneven icing or inconsistent ice shedding at different times can easily cause line breaks and tower collapses; icing on insulator strings can cause frequent flashover accidents; and the dynamic load generated by the galloping of iced conductors can easily damage power equipment such as towers, conductors, and fittings. These accidents and disasters can cause enormous losses and negative impacts on the national economy, society, and people's lives. Therefore, predicting the thickness of icing on transmission lines is of great significance for ensuring the safe and stable operation of the power grid. By predicting the icing thickness and understanding the icing situation of transmission lines in advance, power companies can take timely de-icing measures to carry out anti-icing and maintenance work.

[0003] In the existing technology, due to breakthroughs in artificial intelligence, computing power and other related technologies, deep learning-based ice thickness prediction models have developed rapidly in recent years and have become a highly promising prediction method. Related research mainly focuses on time series prediction and local feature capture to build prediction models, which to a certain extent solves the problems of local feature extraction of ice thickness data and short-term time series prediction accuracy, and provides a foundation for the engineering application of ice prediction.

[0004] However, in actual power transmission line operation scenarios, power transmission lines often cross multiple terrain and meteorological regions. The icing status of each monitoring tower does not exist independently, but is affected by the spatial topological correlation of micro-topography and the regional propagation characteristics of micro-meteorology, resulting in significant global spatial correlation characteristics. Although the current icing thickness prediction model has solved the problem of capturing local features and time series prediction of icing thickness data to some extent, it still fails to solve the problem of insufficient capture of global spatial features of power transmission line icing thickness data and micro-meteorological data. Summary of the Invention

[0005] One objective of this invention is to provide a method for predicting icing on transmission lines based on the fusion of spatiotemporal graph topology and micro-meteorological features. This method constructs a graph model by analyzing the correlation of icing thickness between nodes, the spatial relationships between nodes, and the correlations between meteorological factors. Then, a Geometric Component Network (GCN) model is used to extract features from the graph model, obtaining the spatial features between nodes. The extracted spatial features are then input into an LSTM-SNPS system for temporal feature extraction and icing thickness prediction. This avoids temporal prediction bias caused by missing spatial information, thereby improving the accuracy of icing thickness prediction.

[0006] This objective is achieved using the following technical solution:

[0007] A method for predicting icing on transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features includes:

[0008] Step S1: Obtain ice thickness data and meteorological data at each monitoring point of the transmission line, and preprocess the ice thickness data and meteorological data;

[0009] Step S2: Take each monitoring point as a node, use the preprocessed ice thickness data and meteorological data as node attributes, and construct a graph model by combining the spatial relationship between nodes and the correlation between node attributes.

[0010] Step S3: Extract features from the graph model using the GCN model to obtain node-enhanced features that incorporate spatial association information;

[0011] Step S4: Input the node enhancement features into the LSTM-SNPS model to obtain the predicted value of the icing thickness of the transmission line.

[0012] Meteorological factors such as temperature, humidity, and wind speed are key factors affecting the thickness of icing on transmission lines. Certain meteorological conditions are required for icing to occur. Geographical factors also influence meteorological conditions, indirectly affecting icing thickness. Micro-topography refers to a terrain unit within a mountain range, plain, or basin, shaped by local geomorphological elements such as slope, aspect, surface undulation, gully direction, pass structure, and vegetation cover type and distribution. This micro-topography can significantly differentiate local environmental factors such as airflow, water vapor distribution, and heat exchange. Micro-meteorology refers to the local meteorological state formed within a small area, directly regulated by micro-topographic units, and distinct from the surrounding background meteorology. Slope, aspect, and altitude are key factors influencing meteorological conditions in mountainous areas, directly or indirectly altering spatial temperature characteristics. The meteorological conditions of transmission lines are not isolated but form a spatially continuous meteorological field through atmospheric circulation and topographic barriers. The spatial relationships of this meteorological field directly determine the distribution pattern of icing, thus affecting the icing situation of transmission lines.

[0013] Therefore, this invention proposes a transmission line icing thickness prediction model based on the GCN-LSTM-SNPS model. The GCN model captures spatial dependencies in the data by aggregating node neighborhood information, defines spatial relationships between nodes through a graph structure, and then uses graph convolution operations to allow each node to continuously aggregate its own and its neighbors' features, progressively transmitting and fusing local and global spatial information. This automatically learns and captures the spatial dependencies implicit in icing thickness data and meteorological data, integrating the features of each monitoring point into the spatial relationship information of the entire line. Referencing the real-time situation of surrounding points significantly reduces prediction errors. Simultaneously, by combining the LSTM-SNPS model in a cascaded manner, temporal modeling is performed on the spatial feature sequences extracted by GCN, which can effectively capture long-term dependency patterns of features changing over time. The LSTM-SNPS model combines the temporal modeling advantages of Long Short-Term Memory (LSTM) networks with the spiking neural network (SN) system. The LSTM-SNPS model utilizes a bio-inspired mechanism. When directly facing the high-dimensional spatial enhancement features output by GCN, LSTM is prone to fitting bias and gradient vanishing risks due to excessive redundant information. However, when combined with SNPS, based on the triggering and decay rules of biological neuron pulse transmission, it can dynamically select key features such as critical meteorological conditions for icing and nodes of sudden changes in icing thickness and convert them into pulse signals. This filters out invalid and redundant information, significantly improving the feature utilization efficiency of LSTM, reducing modeling errors caused by irrelevant information, and making the model more consistent with the dynamic evolution logic of actual icing. By capturing spatial dependencies in the data through the GCN model, the LSTM-SNPS model performs temporal modeling on the spatial feature sequences extracted by GCN, which can effectively capture the long-term dependency patterns of features changing over time. This achieves collaborative modeling of spatiotemporal factors, enabling the model to accurately characterize the spatial propagation law and temporal evolution characteristics of icing, and improving the prediction accuracy of icing thickness in complex scenarios.

[0014] Furthermore, step S2 includes the following steps:

[0015] Step S2-1: Calculate the Pearson correlation coefficient between ice thickness and meteorological factors between any two nodes;

[0016] The Pearson correlation coefficient can quantify the degree of linear correlation between two variables. The closer the absolute value is to 1, the stronger the linear correlation. In this step, for each monitoring node of the transmission line, the Pearson correlation coefficient between the ice thickness sequence and meteorological factors between each pair of nodes is calculated. By using statistical methods to explore the linkage strength between ice and meteorological conditions at different locations, the transmission effect of meteorological conditions between nodes can be clarified, providing a scientific basis for the subsequent construction of graph model edges.

[0017] Step S2-2: Set a threshold for the Pearson correlation coefficient, determine nodes with a value greater than the threshold as associated edges, and construct an adjacency matrix;

[0018] By setting a correlation threshold to eliminate redundant node pairs with weak or no correlation, a true spatial linkage relationship is considered to exist between two nodes only when the absolute value of their Pearson correlation coefficient is greater than the preset threshold. This simplifies the graph model topology, improves computational efficiency, and filters out false weak correlations caused by random factors. This allows the graph model topology to focus on real and stable spatial linkage relationships, thereby improving the accuracy of GCN spatial feature extraction.

[0019] Step S2-3: Extract the ice thickness and meteorological features of each node, construct the node feature matrix, and combine the adjacency matrix and the node feature matrix to construct a graph model.

[0020] The node feature matrix is ​​the set of attributes for each monitoring point and is the basic data for GCN to perform feature aggregation. The adjacency matrix is ​​the association rule between nodes, which defines the scope and object of feature aggregation. Combined with the constructed graph model, it retains the attribute features of each node and assigns spatial association attributes to the nodes through the adjacency matrix, providing a data carrier for GCN to capture spatial dependencies.

[0021] By using Pearson correlation coefficient to analyze the correlation between ice thickness and meteorological factors at each node, the spatial linkage between points is quantified. The edges of the graph model are selected based on the correlation threshold to form a topological structure that better fits the physical laws of ice accumulation. This fully utilizes the spatial correlation of ice accumulation, making up for the shortcomings of traditional time series models that cannot capture global spatial features. It has higher prediction accuracy and engineering applicability in complex terrain scenarios.

[0022] Furthermore, the GCN model adopts a two-layer graph convolution structure. Each layer achieves local smooth transmission of node features through the normalization of the adjacency matrix. The number of graph convolution layers directly determines the aggregation range of node features. A single-layer convolution can only aggregate the features of first-order neighbors and can only capture extremely local spatial relationships. Two-layer convolution can be extended to aggregate second-order neighbors, which precisely covers the typical spatial impact scale of icing on transmission lines. The low temperature and high humidity effects in the core icing area usually only transmit to 2-3 adjacent downstream points, which can accurately depict the spatial linkage relationship of icing between adjacent nodes under complex micro-topography in mountainous areas.

[0023] Furthermore, the LSTM-SNPS model adopts the impulse rules of the SNPS model. These impulse rules are used in the gating activation logic of the LSTM-SNPS model. The LSTM-SNPS model replaces the sigmoid / tanh gating mechanism inside the traditional LSTM with the impulse triggering, impulse propagation, and impulse decay rules of SNPS, forming an impulse-driven temporal memory unit. The impulse rules only trigger impulses when there are important signals, such as sudden changes in ice thickness or when meteorological conditions reach the icing threshold. Therefore, the model naturally has dynamic feature filtering capabilities, without the need for additional attention mechanisms, thus improving the efficiency of high-dimensional feature processing. Moreover, since icing has a significant lag, the ice thickness will only change several hours after the weather changes. The impulse delay propagation mechanism of SNPS can automatically learn and match the icing lag pattern by learning the delay step size, which is closer to the actual evolution of icing. This allows for more accurate and efficient capture of the long-term temporal dependence and abrupt change characteristics of ice thickness, significantly improving the accuracy of icing prediction under complex terrain.

[0024] Furthermore, the LSTM-SNPS model includes three nonlinear gates, which are used to control the state update and output intensity of the LSTM-SNPS model. The SNPS pulse signal is a binary discrete output that is triggered or not triggered, making it difficult to quantify the difference in the strength of features. The linear gates can adjust the pulse signal through continuous linear weight coefficients, allowing the model to both select key features through pulses and accurately quantify feature strength through linear weighting, stabilizing the training process and improving the accuracy of icing prediction. The state update of LSTM-SNPS corresponds to the dynamic evolution process of the model's temporal memory core. The state update allows the model to stably retain core features in long-term sequences, improving the reliability of prediction.

[0025] Furthermore, the three nonlinear gates include a reset gate, a consumption gate, and a generation gate:

[0026] The reset gate is used to determine the retention ratio of historical states, reduce interference from redundant information, and determine how many historical states to discard.

[0027] The consumption gate is used to quantify pulse consumption and adjust the influence of historical information on the current state. By quantifying pulse consumption, the consumption gate can assign differentiated consumption weights to historical pulses from different periods, thus avoiding modeling bias caused by overload of historical information.

[0028] The generation gate is used to generate pulse signals and determine the output intensity of the current state. Through a threshold-triggered nonlinear response mechanism, continuous temporal features are transformed into event-driven signals, realizing differentiated encoding of stable backgrounds and icing mutation events, thereby enhancing the dynamic expression capability of sequence features.

[0029] Furthermore, the calculation formula for the reset door is as follows:

[0030] ;

[0031] In the formula, Indicates the reset gate; the current input is The state at the previous moment was ; and This is the weight matrix for the reset gate; This is for resetting the gate bias.

[0032] Furthermore, the calculation formula for the consumption gate is as follows:

[0033] ;

[0034] In the formula, This represents the consumption gate; the current input is... The state at the previous moment was ; and This is the weight matrix of the consumption gate; This is for consuming gate bias.

[0035] Furthermore, the calculation formula for the generation gate is as follows:

[0036] ;

[0037] In the formula, This represents the generation gate; the current input is... The state at the previous moment was ; and This is the weight matrix for generating the gate; To generate gate bias.

[0038] Furthermore, the LSTM-SNPS model determines its output through a reset gate, a consumption gate, and a generation gate. The calculation formula for the model output is as follows:

[0039] ,

[0040] ,

[0041] ;

[0042] In the formula, Indicates the candidate pulse signal; This indicates the current final output of the generation gate; This indicates the updated current state. and This is the weight matrix for the candidate pulse gating coefficients; This is the bias of the candidate pulse gating coefficients.

[0043] The LSTM-SNPS model uses a generation gate to lead candidate pulse generation, a reset gate and a consumption gate to work together to preserve the state and update the consumption, and a generation gate to complete the calculation process of the output gated pulse. It deeply integrates the SNPS pulse triggering, consumption and transmission mechanism with the temporal memory capability of LSTM to achieve event-driven expression of icing temporal features, dynamic memory update and accurate delayed output, which significantly improves the model's ability to capture the icing evolution law and the prediction accuracy.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] 1. The present invention provides a transmission line icing prediction method based on the fusion of spatiotemporal graph topology and micro-meteorological features. By using the neighborhood information of each aggregation node, the method captures the global spatial correlation information of the transmission line, accurately depicts the spatial gradient correlation between nodes and the spatial propagation law of icing, corrects the deviation of the monitoring micro-meteorological values ​​of each transmission line caused by geographical location, and significantly improves the accuracy of icing thickness prediction.

[0046] 2. The transmission line icing prediction method based on the fusion of spatiotemporal graph topology and micro-meteorological features of the present invention deeply integrates the pulse event driving and decay mechanism of the SNPS model and the temporal memory capability of the LSTM model. It utilizes the pulse characteristics of SNPS to mark icing surge events with high pulse values ​​and low pulse values ​​to correspond to stable stages, so that the model can clearly distinguish the feature differences of different icing stages, realize the accurate encoding of non-stationary temporal features, avoid the excessive smoothing of abrupt information by traditional LSTM, and significantly improve the prediction accuracy of icing thickness of key nodes.

[0047] 3. The transmission line icing prediction method based on the fusion of spatiotemporal graph topology and micro-meteorological features of the present invention captures the spatial dependencies in the data through the GCN model, and performs time-series modeling on the spatial feature sequence extracted by GCN using the LSTM-SNPS model. This effectively captures the long-term dependency patterns of features changing over time, realizes the collaborative modeling of spatiotemporal factors, and enables the model to accurately characterize the spatial propagation law and temporal evolution characteristics of icing, thereby improving the prediction accuracy of icing thickness in complex scenarios. Attached Figure Description

[0048] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0049] Figure 1 This is a schematic diagram of the overall steps of a transmission line icing prediction method based on the fusion of spatiotemporal map topology and micro-meteorological features according to the present invention.

[0050] Figure 2 This is a step-by-step schematic diagram of step S2 of the transmission line icing prediction method based on the fusion of spatiotemporal map topology and micro-meteorological features of the present invention.

[0051] Figure 3 This is a schematic diagram of the geographical locations of each node in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram illustrating the correlation of ice thickness at each node in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram illustrating the correlation between ice thickness and meteorological factors at each node in region A in this embodiment of the invention;

[0054] Figure 6 This is a schematic diagram illustrating the correlation between ice thickness and meteorological factors at each node in region B in this embodiment of the invention.

[0055] Figure 7 This is the process of constructing an adjacency matrix using the ice thickness of each node in region A as a feature in this embodiment of the invention;

[0056] Figure 8 This is the network structure of the GCN model in this embodiment of the invention;

[0057] Figure 9 This is the structure of the LSTM-SNPS model in the embodiments of the present invention;

[0058] Figure 10 This is the structure of the GCN-LSTM-SNPS model in the embodiments of the present invention;

[0059] Figure 11 These are radar images of the mean absolute error (MAE) at various nodes and regions for different combinations in embodiments of the present invention;

[0060] Figure 12 These are radar images of the mean absolute error percentage (MAPE) for different combinations at various nodes and regions in embodiments of the present invention.

[0061] Figure 13 This is a radar chart of the root mean square error (RMSE) of different combinations at each node and in each region in the embodiments of the present invention;

[0062] Figure 14 This refers to the goodness of fit (Rfit) of different combinations at each node and in each region in the embodiments of the present invention. 2 Radar chart;

[0063] Figure 15 This is a comparison chart of the mean absolute error (MAE) of icing thickness prediction for a single transmission line and multiple transmission lines in an embodiment of the present invention.

[0064] Figure 16 This is a comparison chart of the mean absolute error percentage (MAPE) of the icing thickness prediction for a single transmission line and multiple transmission lines in an embodiment of the present invention.

[0065] Figure 17 This is a comparison chart of the root mean square error (RMSE) of the icing thickness prediction for a single transmission line and multiple transmission lines in an embodiment of the present invention.

[0066] Figure 18 This refers to the goodness of fit (Ri) between the prediction of icing thickness for a single transmission line and multiple transmission lines in this embodiment of the invention. 2 Comparison chart;

[0067] Figure 19 This is a schematic diagram of the mean absolute error (MAE) of the transmission line icing thickness prediction models GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS in the embodiments of the present invention.

[0068] Figure 20 This is a schematic diagram of the mean absolute error percentage (MAPE) of the transmission line icing thickness prediction models GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS in the embodiments of the present invention.

[0069] Figure 21 This is a schematic diagram of the root mean square error (RMSE) of the transmission line icing thickness prediction models GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS in the embodiments of the present invention.

[0070] Figure 22 The goodness-of-fit (R²) of the transmission line icing thickness prediction models of GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS in the embodiments of the present invention is as follows: 2 ) Schematic diagram;

[0071] Figure 23 This is a comparison chart of the predicted values ​​and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) for node a in this embodiment of the invention.

[0072] Figure 24 This is a comparison chart of the predicted values ​​and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) for node b in this embodiment of the invention.

[0073] Figure 25 This is a comparison chart of the predicted values ​​and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) for node c in this embodiment of the invention.

[0074] Figure 26 This is a comparison chart of the predicted values ​​and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) for node d in this embodiment of the invention.

[0075] Figure 27 This is a comparison chart of the predicted and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) for node e in this embodiment of the invention. Detailed Implementation

[0076] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0077] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0078] Example 1

[0079] A method for predicting icing on transmission lines based on the fusion of spatiotemporal graph topology and micrometeorological features, such as... Figure 1 As shown, it includes:

[0080] Step S1: Obtain ice thickness data and meteorological data at each monitoring point of the transmission line, and preprocess the ice thickness data and meteorological data;

[0081] The meteorological data includes temperature, relative humidity, wind speed, and atmospheric pressure at each monitoring point.

[0082] In some embodiments, a normalization method is used to preprocess the ice thickness data and meteorological data.

[0083] Because historical data on icing thickness of transmission lines and meteorological monitoring data are discontinuous and contain missing or abnormal data, this embodiment uses 25-minute linear interpolation to complete all monitoring data within the statistical timeframe. Finally, the data is normalized using the formula shown.

[0084] ;

[0085] In the formula, This is the initial data; and They are respectively The minimum and maximum values ​​in the column; This is the normalized data.

[0086] Step S2: Take each monitoring point as a node, use the preprocessed ice thickness data and meteorological data as node attributes, and construct a graph model by combining the spatial relationship between nodes and the correlation between node attributes.

[0087] like Figure 2 As shown, step S2 includes the following steps:

[0088] Step S2-1: Calculate the Pearson correlation coefficient between ice thickness and meteorological factors between any two nodes;

[0089] This embodiment uses historical icing thickness data and historical meteorological data of five transmission lines in two typical micro-topographic and micro-meteorological regions in China. The two regions are labeled A and B. Region A contains three transmission lines, denoted as a, b, and c. The experimental data includes historical meteorological monitoring data and icing thickness data from January 19, 2025 to February 19, 2025. Region B contains two transmission lines, denoted as d and e. The experimental data includes historical icing thickness data and meteorological monitoring data from December 6, 2024 to January 6, 2025.

[0090] The geographical locations of each transmission line are as follows: Figure 3 As shown, both the horizontal and vertical axes represent location distances, used to quantify the spatial location and relative positional relationship of each monitoring point.

[0091] Calculate the Pearson correlation coefficient between icing thickness and meteorological factors between transmission line nodes in regions A and B, such as... Figure 4The diagram shows the correlation between icing thickness among nodes, used to demonstrate the spatial correlation strength or feature similarity between transmission line monitoring nodes. The horizontal and vertical axes represent the node numbers of the transmission line monitoring nodes, namely a, b, c, d, e. Red represents high values ​​(close to 1), indicating a stronger positive correlation between the corresponding two nodes, while blue represents low values ​​(close to -1), indicating a stronger negative correlation between the corresponding two nodes.

[0092] like Figure 5 The diagram shows the correlation between icing thickness and meteorological factors on various transmission lines in region A. It visually demonstrates the strength and direction of linear correlations between different features. The horizontal and vertical axes represent feature types, including: temperature (a-temperature), temperature (b-temperature), and temperature (c-temperature) at node a; humidity (a-humidity), humidity (b-humidity), and humidity (c-humidity) at node a; wind speed (a-wind_speed), wind speed (b-wind_speed), and wind speed (c-wind_speed) at node b; and atmospheric pressure (a-pressure), atmospheric pressure (b-pressure), and atmospheric pressure (c-pressure) at node a. The feature lists on both the horizontal and vertical axes are identical, used to calculate the Pearson correlation coefficient pairwise.

[0093] like Figure 6 The diagram shows the correlation between icing thickness and meteorological factors on various transmission lines in region B. It visually demonstrates the strength and direction of linear correlations between different features. The horizontal and vertical axes represent feature types, including temperature (d-temperature) and temperature (e-temperature) at node d, humidity (d-humidity) and humidity (e-humidity) at node d and node e, wind speed (d-wind_speed) and wind speed (e-wind_speed) at node d and atmospheric pressure (d-pressure) and atmospheric pressure (e-pressure) at node d and node e. The feature lists on the horizontal and vertical axes are identical, used to calculate the Pearson correlation coefficient pairwise.

[0094] Step S2-2: Set a threshold for the Pearson correlation coefficient, determine nodes with a value greater than the threshold as associated edges, and construct an adjacency matrix;

[0095] The process of constructing the adjacency matrix is ​​as follows: Figure 7 As shown, with Figure 4Taking the adjacency matrix with ice thickness as a feature in region A as an example, the Pearson correlation coefficient threshold is set to ±0.8. That is, when the Pearson correlation coefficient of ice thickness between two nodes is ≥ 0.8 or ≤ -0.8, it is determined that there is a strong correlation between the two nodes, and it is marked as 1 in the adjacency matrix (indicating the existence of a related edge). When the correlation coefficient is in the range of (-0.8, 0.8), it is determined that there is a weak correlation between the two nodes, and it is marked as 0 in the adjacency matrix (indicating no related edge). Based on this... Figure 5 Region A and Figure 6 The diagram illustrating the correlation between ice thickness and meteorological factors at each node in region B is used to construct adjacency matrices corresponding to ice thickness and meteorological factors, thus completing the construction of adjacency matrices with multiple feature dimensions.

[0096] Step S2-3: Extract the ice thickness and meteorological features of each node, construct the node feature matrix, and combine the adjacency matrix and the node feature matrix to construct a graph model.

[0097] The ice thickness feature and meteorological factor feature of each node are extracted separately. For each node, the measured time series value of ice thickness within a fixed time window is extracted as the ice thickness feature, and the measured values ​​of core meteorological indicators such as temperature, relative humidity, wind speed, and atmospheric pressure, which are synchronized with the ice thickness, are extracted as the meteorological factor features. Assuming there are N transmission line monitoring nodes, the fused feature vectors of all nodes are arranged column-wise and row-wise to generate a node feature matrix X with dimensions N×D, where N is the total number of nodes and D is the sum of the dimensions of the ice thickness time series feature and the meteorological factor feature. For each transmission line node i, its ice thickness time series feature vector and the multi-dimensional meteorological factor feature vector are concatenated row-wise to generate the fused feature vector x of that node. i The node feature matrix X=[x1,x2,...,x] N ] T In this embodiment, N=5, and D is the total number of features of icing thickness and meteorological factors. The meteorological features include indicators such as temperature, relative humidity, wind speed, and atmospheric pressure. This matrix fully represents the icing and meteorological multidimensional feature information of all nodes. The transmission line monitoring nodes are taken as the node set V, and the adjacency matrix A constructed based on the Pearson correlation coefficient of icing thickness and meteorological factors determines the edge set E between nodes. The icing thickness and meteorological factor features extracted from each node are constructed into a node feature matrix X. The node set V, the edge set E, the node feature matrix X, and the adjacency matrix A together constitute an attributed graph model G=(V,E,X,A). GCN realizes the spatial aggregation of node features through this graph model. Relying on the topological modeling logic of graph neural networks, combined with the node spatial association information of the adjacency matrix and the self-attribute information of the node feature matrix, the graph model for transmission line icing prediction is completed.

[0098] Step S3: Extract features from the graph model using the GCN model to obtain node-enhanced features that incorporate spatial association information;

[0099] In some embodiments, the GCN model adopts a two-layer graph convolution structure, and each layer achieves local smooth transmission of node features through the normalization of the adjacency matrix.

[0100] GCN is a graph-based deep learning model that primarily extracts node features by performing convolution operations on the graph, and then uses these features for tasks such as classification and clustering. Unlike traditional convolutional neural networks, GCN's convolutional layers use adjacency matrices, degree matrices, and other methods to define the graph structure, thereby enabling convolution operations on the features of nodes in the graph.

[0101] If we set up a diagram ,in It is a set of nodes. If the set is an edge set, then the single-layer iteration formula for GCN is:

[0102] ;

[0103] In the formula, , For the image The edge information in the adjacency matrix, for An identity matrix of order 1. The number of nodes; Then it is a diagram The adjacency matrix contains the feature information of nodes and the feature information of edges, which is used to define the connection relationship between nodes and control the scope of information propagation; yes The degree matrix is ​​a diagonal matrix; It is the first Layer node feature vectors; It is the first Layer weight matrix; Represents the number of layers in the network; It is an activation function.

[0104] The feature extraction process in GCN involves using information about the nodes themselves and the edges between them to extract features from the graph structure. The GCN network structure is as follows: Figure 8 As shown, each input node X contains information about its own node. The updated node information Y is obtained by performing a convolution operation using the information of its own node and its edges. For example, Y3 is the updated node obtained by convolving the information of the edges between X3 and X1, X2, X4, and X5, as well as the node information of X3 itself.

[0105] Where X is the initial feature representation of each transmission line node, corresponding to the first feature in the GCN formula. Layer node feature vector The X of a single node includes the icing thickness characteristics and meteorological factors of that transmission line node itself. The X of all nodes are organized in matrix form, which is the node feature matrix.

[0106] Y is the feature representation of each transmission line node obtained after GCN convolution operation, corresponding to the output after l-layer iteration in the GCN formula. The GCN model mines the spatial association between a node and its neighboring nodes, such as X1, X2, X4, and X5, of X3 through the adjacency matrix. It aggregates the information of node X3 itself with the information of its neighboring nodes X1, X2, X4, and X5 according to the spatial association degree, and then transforms them through the activation function and weight matrix to obtain the node enhancement feature Y3. Similarly, for nodes X1, X2, X4, and X5, the information of each node itself and the information of its neighboring nodes are aggregated according to the spatial association degree, and then transformed through the activation function and weight matrix to obtain the corresponding node enhancement features Y1, Y2, Y4, and Y5.

[0107] Step S4: Input the node enhancement features into the LSTM-SNPS model to obtain the predicted value of the icing thickness of the transmission line.

[0108] In some embodiments, the LSTM-SNPS model adopts the impulse rules of the SNPS model. The impulse rules are used for the gating activation logic of the LSTM-SNPS model. LSTM-SNPS is a novel long short-term memory network inspired by the Spiking Neural PSystem (SNPS) mechanism. It replaces the sigmoid or tanh gating of traditional LSTM with the impulse rules of SNPS, inheriting the recurrent structure and time series modeling capabilities of LSTM.

[0109] In some embodiments, such as Figure 9 As shown, The LSTM-SNPS model, which represents element-wise multiplication, includes three nonlinear gates that control the state update and output intensity of the LSTM-SNPS model.

[0110] In some embodiments, the three nonlinear gates include a reset gate, a consumption gate, and a generation gate:

[0111] Let the current input be The state at the previous moment was The calculation formula is as follows:

[0112] ,

[0113] ;

[0114] In the formula, e is the base of the natural logarithm, approximately equal to 2.71828. For input variables, The output value is controlled between (−1, 1) and is used to generate candidate states or outputs. Map the input to (0,1) to generate the gating coefficients (e.g., Candidate pulse signals), and the retention or discarding of control information.

[0115] The reset gate is used to determine the retention ratio of historical states;

[0116] The calculation formula for the reset gate is:

[0117] ;

[0118] In the formula, Indicates the reset gate; the current input is The state at the previous moment was ; and This is the weight matrix for the reset gate; This is for resetting the gate bias.

[0119] The consumption gate is used to quantify the pulse consumption and adjust the influence of historical information on the current state.

[0120] The calculation formula for the consumption gate is as follows:

[0121] ;

[0122] In the formula, This represents the consumption gate; the current input is... The state at the previous moment was ; and This is the weight matrix of the consumption gate; This is for consuming gate bias.

[0123] The generation gate is used to generate a pulse signal and determine the output intensity of the current state.

[0124] The formula for calculating the generation gate is:

[0125] ;

[0126] In the formula, This represents the generation gate; the current input is... The state at the previous moment was ; and This is the weight matrix for generating the gate; To generate gate bias.

[0127] The LSTM-SNPS model determines the updated state and model output through a reset gate, an exhaustion gate, and a generation gate. The calculation formulas for the updated state and model output are as follows:

[0128] ,

[0129] ;

[0130] In the formula, Indicates the candidate pulse signal; This indicates the current final output of the generation gate; Indicates the updated current state, symbol This indicates element-wise multiplication.

[0131] Among them, candidate pulse signal The calculation formula is:

[0132] ;

[0133] In the formula, the current input is The state at the previous moment was ; and This is the weight matrix for the candidate pulse gating coefficients; This is the bias of the candidate pulse gating coefficients.

[0134] Through the reset gate and status update It effectively memorizes long-term icing trends while avoiding the gradient vanishing problem, and also provides candidate pulse signals. It can provide a pulsed response to critical events such as sudden increases and decreases in icing, avoiding the over-smoothing of abrupt signals by traditional LSTMs, and its state update formula... Represents the continuation of a historically frozen state. Representing the accumulation or melting of new icing under current meteorological conditions, the state update formula adapts to the lag effect of icing formation by accumulating historical states and current meteorological conditions, and is more consistent with the actual icing thickness accumulation process.

[0135] In some embodiments, the GCN model and the LSTM-SNPS model are combined in a cascaded manner, such as... Figure 10 The structure of the GCN-LSTM-SNPS model is shown. GCN is used to extract features from the graph model to obtain the spatial features between nodes. The extracted spatial features are then input into LSTM-SNPS for temporal feature extraction and to complete the prediction of icing thickness.

[0136] Example 2

[0137] Based on Example 1, in order to verify that the GCN-LSTM-SNPS model for predicting the icing thickness of transmission lines can capture more comprehensive icing thickness trend features by utilizing the spatial relationship between different transmission lines, the icing thickness predictions for multiple transmission lines and a single transmission line were compared in regions A and B respectively.

[0138] The GCN-LSTM-SNPS prediction model consists of two parts: GCN feature extraction and LSTM-SNPS time series prediction. First, in the GCN feature extraction part, a two-layer GCN structure is used with ReLU activation. Second, in the LSTM-SNPS time series prediction part, the historical time series length is set to 36, and the activation function is tanh. To verify the accuracy of the prediction results, the experimental data is divided at a ratio of 0.7, i.e., 70% of the experimental data is used as the training set and 30% as the validation set.

[0139] Therefore, during model training, this embodiment sets the core parameters for training the GCN-LSTM-SNPS model as follows: the correlation coefficient for constructing the adjacency matrix in region A is set to greater than ±0.8. Increasing the adjacency matrix construction condition at this level excludes moderately correlated but non-critical adjacency information. The correlation coefficient for constructing the adjacency matrix in region B is set to greater than ±0.7. Appropriately decreasing the adjacency matrix construction condition at this level retains more moderately correlated adjacency information. The remaining parameters are set as follows: the learning rate is set to 0.01, the training batch size is set to 64, the maximum number of iterations is set to 100, and to ensure experimental reproducibility, the number of random seeds is fixed at 42.

[0140] This embodiment uses four evaluation criteria: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Error (RMSE), and Goodness of Fit (R²). 2 The calculation formulas are as follows:

[0141] ,

[0142] ,

[0143] ,

[0144] ;

[0145] In the formula, The actual value; This is a predicted value; The average value of the predicted data; This represents the amount of data to be predicted.

[0146] In this embodiment, four sets of meteorological factors are combined with icing thickness to predict icing thickness using the GCN-LSTM-SNPS model. The optimal combination of meteorological factors and icing thickness is selected by comparing and evaluating the errors. For ease of description, the combination of temperature, humidity, and wind speed is abbreviated as THW, the combination of temperature, humidity, and air pressure is abbreviated as THP, and similarly, the combination of humidity, wind speed, and air pressure is abbreviated as HWP, and the combination of temperature, humidity, wind speed, and air pressure is abbreviated as THWP.

[0147] Different meteorological factors and icing thickness were combined to obtain the icing thickness prediction results of the GCN-LSTM-SNPS model for transmission lines with different combinations. The evaluation error is as follows: Figure 11 to 14 As shown, Figure 11 This is a radar chart showing the mean absolute error (MAE) of different combinations of meteorological factors at nodes a, b, c, d, e, and in regions A and B. Figure 12 A radar chart (MAPE) showing the percentage of mean absolute error for different combinations of meteorological factors at nodes a, b, c, d, e, and in regions A and B. Figure 13 Radar charts showing the root mean square error (RMSE) of different combinations of meteorological factors at nodes a, b, c, d, e, and in regions A and B. Figure 14 The R-value represents the goodness of fit (R²) of different combinations of meteorological factors at nodes a, b, c, d, e, and in regions A and B. 2 (Radar image)

[0148] The evaluation errors of regions A and B show that the meteorological factor combinations THW and THP have the highest prediction accuracy. Subsequent comparative experiments will only analyze the prediction of icing thickness under the meteorological combinations of THW or THP.

[0149] like Figures 15 to 18 As shown, this is used to evaluate the prediction performance of different models at different nodes. Figures 15 to 18 The x-coordinates represent different nodes a, b, c, d, and e. Figures 15 to 18 The vertical axis represents the mean absolute error (MAE), mean absolute error percentage (MAPE), root mean square error (RMSE), and goodness of fit (R²). 2The different models are THW-single-line, THW-multi-line, THP-single-line, and THP-multi-line. Single-line icing thickness prediction refers to inputting only the meteorological and icing thickness characteristics of one line for icing prediction, and outputting only the predicted icing thickness value for that line. Multi-line icing thickness prediction refers to inputting the icing thickness characteristics and meteorological icing predictions of multiple transmission lines, and simultaneously outputting the predicted icing thickness values ​​for all multiple transmission lines.

[0150] like Figures 15 to 18 As shown, the GCN-LSTM-SNPS model for predicting icing thickness of transmission lines significantly improves the accuracy of icing thickness prediction for each line by capturing the spatial relationships between multiple transmission lines. Specifically, in predicting the icing thickness of transmission line b in region A, the goodness of fit is negative in both the THW and THP weather combinations, with MAE and MAPE exceeding 0.4 mm and 0.36 respectively, indicating a prediction accuracy below 65%. However, in predicting the icing thickness of multiple transmission lines, the accuracy is significantly improved. The errors MAE and MAPE of transmission line b are reduced to below 0.22 mm and 0.11 respectively, indicating a prediction accuracy of 89%. Furthermore, the evaluation error of transmission line d in region B within the THP meteorological combination also shows a significant improvement in all evaluation errors compared to the prediction of icing thickness for a single transmission line. The goodness of fit changes from negative to positive, and MAPE decreases from 1.0124 to 0.0626, indicating an improvement in prediction accuracy to 93%. MAE and RMSE also decrease from 1.1837 mm and 1.1936 mm to 0.0702 mm and 0.0775 mm, respectively. In addition, the other three transmission lines a, c, and e also demonstrate that the prediction errors for icing thickness from multiple transmission lines are comprehensively lower than those from a single transmission line. Therefore, the GCN-LSTM-SNPS transmission line icing thickness prediction model improves prediction accuracy by capturing the global spatial relationship between icing thicknesses of transmission lines and correcting for deviations in micro-meteorological data monitored by each transmission line due to geographical location.

[0151] Example 3

[0152] Based on Examples 1 and 2, the transmission line icing thickness prediction model of this example is compared with other common transmission line icing thickness prediction models in a comparative experiment.

[0153] The comparative experiment compared the results of five transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, and GCN-LSTM) with the GCN-LSTM-SNPS transmission line icing thickness prediction model in this embodiment under the THW meteorological combination.

[0154] Five transmission line icing thickness prediction models—GCN, LSTM, BPNN, CNN-LSTM, and GCN-LSTM—and the GCN-LSTM-SNPS transmission line icing thickness prediction model in this embodiment were used to predict icing thickness, respectively. Figures 19 to 22 This diagram illustrates the error evaluation between five transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM) and the GCN-LSTM-SNPS transmission line icing thickness prediction model in this embodiment. The horizontal axis represents different transmission line nodes (a, b, c, d, e), and the vertical axis represents the mean absolute error (MAE), mean absolute error percentage (MAPE), root mean square error (RMSE), and goodness of fit (R²). 2 ).

[0155] like Figures 19 to 22 As shown, the GCN-LSTM-SNPS combined prediction model demonstrates significantly better performance than the single prediction models GCN and LSTM in predicting icing thickness of transmission lines. In the single GCN transmission line icing thickness prediction model, the evaluation error value of transmission line d in region B is too high, with MAE, MAPE, and RMSE reaching 1.2581 mm, 1.0751 mm, and 1.2674 mm, respectively, and the goodness of fit is 0.6761, which is far from meeting the prediction requirements. This indicates that although GCN can capture the spatial feature information of icing thickness between transmission lines, it lacks the ability to capture the dynamic formation characteristics of icing thickness of transmission lines.

[0156] In the prediction of icing thickness of a single LSTM transmission line, although the evaluation errors of the d transmission line in region B were reduced by 1.2131 mm, 1.0241 mm, and 1.1918 mm respectively compared with the single GCN transmission line icing thickness prediction model, the goodness of fit only increased by 0.0764. This indicates that although LSTM can perform time-series prediction of icing thickness, it cannot integrate the spatial feature information between transmission lines.

[0157] In the GCN-LSTM-SNPS transmission line icing thickness prediction model, the goodness of fit for transmission line d in region B reached 0.9058. Compared with the single GCN and LSTM prediction models, the goodness of fit of the GCN-LSTM-SNPS model was 0.2297 and 0.1533 higher, respectively. Furthermore, although the error reduction of the GCN-LSTM-SNPS transmission line icing thickness prediction model was not significant compared with the single LSTM model, it was significantly lower than the error of the single GCN model. MAE, MAPE, and RMSE decreased by 1.2192 mm, 1.0401 mm, and 1.2204 mm, respectively. All error indices of the GCN-LSTM-SNPS model were lower than those of the GCN-LSTM, CNN-LSTM, and BPNN models, and the goodness of fit of the GCN-LSTM-SNPS model was also higher than that of the GCN-LSTM, CNN-LSTM, and BPNN models.

[0158] In summary, the GCN-LSTM-SNPS transmission line icing thickness prediction model effectively solves the problem of single models neglecting the interaction between spatial relationships and temporal characteristics by capturing the spatial features of transmission lines through GCN and the temporal features of icing thickness through LSTM.

[0159] The predicted and actual icing thickness values ​​of five transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, and GCN-LSTM) and the GCN-LSTM-SNPS transmission line icing thickness prediction model in this embodiment are compared. Figures 23 to 27 The figures shown are comparisons between the predicted and monitored values ​​of six transmission line icing thickness prediction models (GCN, LSTM, BPNN, CNN-LSTM, GCN-LSTM, and GCN-LSTM-SNPS) at nodes (a, b, c, d, e), where the horizontal axis represents the time axis and the vertical axis represents the icing thickness.

[0160] Depend on Figure 25 It can be seen that in the icing prediction of transmission line c in region A, the GCN-LSTM model shows a similar trend to the actual value, but the numerical difference is large. When using the CNN-LSTM model to predict the icing thickness of transmission lines, the prediction results are not ideal. The goodness of fit for transmission lines d and e in region B is negative, and the evaluation errors MAPE are as high as 0.1584 and 0.1796, respectively. In addition, the predicted value of transmission line d is all 0, which means it is an invalid prediction. At the same time, for the remaining transmission lines a, b, and c in region A, the evaluation errors of the CNN-LSTM model are all lower than those of the GCN-LSTM-SNPS model.

[0161] When using the BPNN model to predict the icing thickness of transmission lines, the goodness of fit for transmission line c in region A reached as high as 0.9212. However, its evaluation errors (MAE) were 0.4616 mm and MAPE were 0.3588, which are 2.5-6 times higher than those of the GCN-LSTM-SNPS model. Meanwhile, for transmission lines a and b in region A and transmission lines d and e in region B, the evaluation errors of the BPNN model were all lower than those of the GCN-LSTM-SNPS model. The predicted trend of icing thickness of transmission lines using the BPNN model was highly similar to the actual value, but a significant error still existed between the predicted and actual values.

[0162] In the experimental results of GCN-LSTM-SNPS transmission line icing thickness prediction, although the accuracy of transmission line b in region A was not as high as the other four transmission lines, it had a lower evaluation error and higher goodness of fit than the BPNN and CNN-LSTM transmission line icing thickness prediction models. The remaining transmission lines, including transmission lines a and c in region A and transmission lines d and e in region B, had evaluation errors MAE and MAPE of less than 0.1 mm and 0.06 mm, respectively, meaning their accuracy was higher than 94%. Furthermore, the GCN-LSTM-SNPS transmission line icing thickness prediction results not only showed a high degree of consistency with the actual values ​​in terms of prediction trends, but also exhibited small errors between the predicted and actual values.

[0163] In summary, the GCN-LSTM-SNPS transmission line icing thickness prediction model shows improved accuracy and the highest goodness of fit compared to the GCN, LSTM, GCN-LSTM, BPNN, and CNN-LSTM prediction models, resulting in the best prediction performance.

[0164] This embodiment presents a GCN-LSTM-SNPS transmission line icing thickness prediction model. This model uses the Pearson correlation of the GCN model to construct a graph model and extracts the spatial information features between each transmission line. With the help of LSTM-SNPS time-series dynamic prediction, it further explores the spatiotemporal features of transmission line icing thickness, ultimately improving the accuracy of icing thickness prediction.

[0165] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0166] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting icing on transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features, characterized in that, include: Step S1: Obtain ice thickness data and meteorological data at each monitoring point of the transmission line, and preprocess the ice thickness data and meteorological data; Step S2: Take each monitoring point as a node, use the preprocessed ice thickness data and meteorological data as node attributes, and construct a graph model by combining the spatial relationship between nodes and the correlation between node attributes. Step S3: Extract features from the graph model using the GCN model to obtain node-enhanced features that incorporate spatial association information; Step S4: Input the node enhancement features into the LSTM-SNPS model to obtain the predicted value of the icing thickness of the transmission line.

2. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 1, characterized in that, Step S2 includes the following steps: Step S2-1: Calculate the Pearson correlation coefficient between ice thickness and meteorological factors between any two nodes; Step S2-2: Set a threshold for the Pearson correlation coefficient, determine nodes with a value greater than the threshold as associated edges, and construct an adjacency matrix; Step S2-3: Extract the ice thickness and meteorological features of each node, construct the node feature matrix, and combine the adjacency matrix and the node feature matrix to construct a graph model.

3. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 1, characterized in that, The GCN model employs a two-layer graph convolution structure.

4. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 1, characterized in that, The LSTM-SNPS model uses the impulse rules of the SNPS model, and these impulse rules are used for the gating activation logic of the LSTM-SNPS model.

5. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 1, characterized in that, The LSTM-SNPS model includes three nonlinear gates, which are used to control the state update and output strength of the LSTM-SNPS model.

6. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 5, characterized in that, The three nonlinear gates include a reset gate, a consumption gate, and a generation gate: The reset gate is used to determine the retention ratio of historical states; The consumption gate is used to quantify the pulse consumption and adjust the influence of historical information on the current state. The generation gate is used to generate a pulse signal and determine the output intensity of the current state.

7. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 6, characterized in that, The calculation formula for the reset gate is: ； In the formula, Indicates the reset gate; the current input is ; The state at the previous moment was ; and Here is the weight matrix for the reset gate; This is for resetting the gate bias.

8. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 6, characterized in that, The calculation formula for the consumption gate is as follows: ； In the formula, This represents the consumption gate; the current input is... ; The state at the previous moment was ; and This is the weight matrix of the consumption gate; This is for consuming gate bias.

9. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 6, characterized in that, The formula for calculating the generation gate is: ； In the formula, This represents the generation gate; the current input is... ; The state at the previous moment was ; and This is the weight matrix for generating the gate; To generate gate bias.

10. The method for predicting icing of transmission lines based on the fusion of spatiotemporal map topology and micro-meteorological features according to claim 6, characterized in that, The LSTM-SNPS model determines the updated state and model output through a reset gate, an exhaustion gate, and a generation gate. The calculation formulas for the updated state and model output are as follows: ， ， ； In the formula, Indicates the candidate pulse signal; This indicates the current final output of the generation gate; This indicates the updated current state. and This is the weight matrix for the candidate pulse gating coefficients; This is the bias of the candidate pulse gating coefficients.

Images