Icing thickness long-time sequence prediction method, system, equipment and medium

Abstract

The invention discloses an icing thickness long-time sequence prediction method, system and device and a medium, and the method comprises the following steps: obtaining an icing state data set, carrying out the preprocessing of the icing state data set, and obtaining a standardized data set; constructing a time-aware neural network model according to the standardized data set; based on the time-varying data and the fixed data in the standardized data set, constructing a multi-channel fusion network by using a time-aware neural network model to obtain a fusion prediction model; training the fusion prediction model through a training strategy to obtain a trained prediction model; and predicting the icing thickness by using the trained prediction model. According to the method, the time-sensing LSTM network is adopted to process the problem of time discontinuity of SAR inversion data, and the time dynamic characteristics of icing thickness data can be accurately modeled by introducing a time interval weight mechanism.

Description

Technical Field

[0001] This invention relates to the field of power system disaster prevention technology, and in particular to a method, system, equipment and medium for long-term prediction of icing thickness. Background Technology

[0002] Currently, the main methods for monitoring ice cover include optical remote sensing and microwave remote sensing (active and passive). Optical remote sensing is susceptible to solar radiation and cloud cover, and it cannot explore the internal structure of the ice layer, making it unsuitable for studies on ice cover thickness, snow water equivalent, and other ice layer structure parameters. Passive microwave remote sensing can penetrate the surface of ice and snow to obtain information about the ice and snow interior, making it the most common data source for long-term monitoring and research on ice and snow parameters such as ice cover thickness and snow water equivalent. However, existing passive microwave remote sensing images have low spatial resolution, which cannot meet the requirements for high-resolution, high-precision ice cover parameter studies. Compared to optical and passive microwave remote sensing, although active microwave remote sensing has all-weather, all-day Earth observation capabilities, is unaffected by cloud cover and can penetrate the ice and snow surface, and has higher spatial resolution, the images acquired are discontinuous due to factors such as satellite revisit cycles and weather conditions, thus preventing the inversion of continuous long-term ice cover thickness series.

[0003] In recent years, the intersection of remote sensing and deep learning has become a research hotspot. For the prediction of long-term series problems, RNN (Recurrent Neural Network) network structures are usually used. The most commonly used improved variant of RNN is LSTM (Long Short-Term Memory Network), which effectively solves the gradient vanishing and gradient exploding problems of traditional RNN networks and can achieve "long-term memory". However, both RNN and LSTM assume that the data is sampled at equal intervals, ignoring the time interval information of the data. This can lead to insufficient modeling of the relationship between events over long time intervals and overemphasis on events over short time intervals. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method, system, device and medium for predicting long-term ice thickness to solve the problem that existing methods ignore the time interval information of the data, which leads to insufficient modeling of event relationships over long time intervals.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting long-term ice thickness, comprising the following steps:

[0008] Obtain the icing status dataset, and preprocess the icing status dataset to obtain a standardized dataset;

[0009] Construct a time-aware neural network model based on the standardized dataset;

[0010] Based on the time-varying and fixed data in the standardized dataset, a multi-channel fusion network is constructed using the time-aware neural network model to obtain a fusion prediction model.

[0011] The fusion prediction model is trained using a training strategy to obtain a trained prediction model.

[0012] The trained prediction model is used to predict the thickness of the ice layer.

[0013] As a preferred embodiment of the long-term series prediction method for icing thickness described in this invention, the step of preprocessing the icing state dataset includes:

[0014] The asynchronously sampled data is unified to a preset time grid using interpolation methods;

[0015] Standardization methods were used to standardize the sequence data.

[0016] The icing status dataset includes historical icing thickness data, meteorological data, and line characteristic data. The meteorological data and historical icing thickness data are time-varying data, while the line characteristic data are fixed data.

[0017] As a preferred embodiment of the long-term series prediction method for icing thickness described in this invention, the construction steps of the time-aware neural network model include:

[0018] The short-term memory component is adjusted by weighting the time intervals between adjacent data to obtain a reduced short-term memory.

[0019] The long-term memory is combined with the reduced short-term memory to form a complete memory component;

[0020] A standard gating structure is constructed based on the complete memory components.

[0021] The beneficial effects of this preferred technical solution are: it constructs a neural network model that can handle data with irregular time intervals, and through a time weight adjustment mechanism, enables the neural network model to capture the temporal dynamic characteristics of ice thickness data, thereby improving the accuracy of long-term series prediction.

[0022] As a preferred embodiment of the long-term series prediction method for icing thickness described in this invention, the step of constructing the multi-channel fusion network includes:

[0023] Time-varying data is input into a time-series channel composed of a time-aware neural network for processing to obtain time-series features;

[0024] Fixed data is input into a static channel composed of a fully connected neural network for processing to obtain static features;

[0025] The temporal and static features are concatenated and fused, and a prediction result is generated through a fully connected layer.

[0026] The beneficial effects of this preferred technical solution are: to achieve effective integration of dynamic meteorological information and static line characteristics; and through a dual-channel parallel processing architecture, to capture both the dynamic correlation of time-varying factors and encode the physical influence of fixed attributes.

[0027] As a preferred embodiment of the long-term series prediction method for icing thickness described in this invention, the step of training the fusion prediction model includes:

[0028] The first stage involves training using a basic loss function until the fusion prediction model learns the basic data patterns.

[0029] The second stage incorporates a physical rule loss function to force the learning of physical laws;

[0030] The third stage incorporates adversarial training, and finally outputs the trained prediction model.

[0031] The basic loss function is the mean square error between the predicted value and the true value, and the physical rule loss function is used to penalize predictions that violate physical laws.

[0032] The beneficial effects of this preferred technical solution are: by adopting a progressive, phased training strategy, the model is ensured to follow physical constraints based on learning the patterns in the data, thus avoiding prediction results that violate common sense in physics.

[0033] As a preferred embodiment of the long-term icing thickness prediction method of the present invention, the adversarial training step includes:

[0034] Add Gaussian noise to the input data and construct data samples under extreme weather conditions;

[0035] Data with Gaussian noise and data samples under extreme weather conditions are input into the fusion prediction model for training;

[0036] Update model parameters using the backpropagation algorithm;

[0037] Adjust the intensity and frequency of adversarial training to improve the robustness of the fusion prediction model.

[0038] As a preferred embodiment of the long-term series prediction method for icing thickness described in this invention, the step of predicting icing thickness using a trained prediction model includes:

[0039] Acquire real-time meteorological data and line characteristic parameters, and standardize the real-time data using the same preprocessing method as in the training phase;

[0040] The preprocessed real-time meteorological data and historical icing thickness data are input into the time series channel and the line characteristic parameters are input into the static channel. Forward propagation calculation is performed through the trained multi-channel fusion network.

[0041] Obtain the predicted value of ice thickness, perform denormalization on the prediction result to restore the original dimension, and conduct risk assessment and early warning judgment based on the preset safety threshold.

[0042] The prediction results will be output in time series format.

[0043] Secondly, the present invention provides a long-term series prediction system for icing thickness, comprising a data processing module, a model building module, a model training module, and a prediction module:

[0044] The data processing module is responsible for collecting the icing status dataset and performing preprocessing operations on the acquired icing status dataset.

[0045] The model building module designs and builds a time-aware neural network model based on the standardized dataset;

[0046] The model training module formulates and executes a training strategy to train the constructed fusion prediction model;

[0047] The prediction module predicts the icing thickness based on the input data.

[0048] Thirdly, the present invention provides an electronic device, comprising:

[0049] Memory and processor;

[0050] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the long-term series prediction method for ice thickness.

[0051] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the long-term series prediction method for icing thickness.

[0052] Compared with existing technologies, the advantages of this invention are as follows: First, it employs a time-aware LSTM network to address the temporal discontinuity of SAR inversion data. By introducing a time interval weighting mechanism, it can accurately model the temporal dynamic characteristics of icing thickness data, significantly improving the prediction accuracy of long-term series compared to traditional LSTM networks. Second, the dual-channel fusion network architecture organically combines dynamic meteorological factors with static line characteristics. The time-series channel captures the dynamic correlation between meteorology and icing, while the static channel encodes the physical properties of the conductor. Their collaborative operation greatly improves prediction accuracy. Simultaneously, the multi-source data fusion strategy fully utilizes the high-resolution SAR inversion results and rich meteorological auxiliary information, overcoming the limitations of a single data source.

[0053] This invention innovatively combines physical constraints with deep learning. Through a phased training strategy and a composite loss function design, it ensures that the prediction results conform to both data patterns and physical common sense. Adversarial training further enhances the model's robustness to noise and extreme weather, enabling the system to maintain stable and reliable prediction performance even in real-world complex environments. Overall, this invention provides a high-precision, high-reliability solution for predicting icing thickness in power systems, enabling accurate risk identification and early warning, significantly improving the safety of power grid operation, and reducing the risk of power outages and economic losses caused by icing. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is an overall structural diagram of the long-term ice thickness prediction method according to an embodiment of the present invention.

[0056] Figure 2 This is a T-LSTM cell structure diagram of the long-term ice thickness prediction method according to an embodiment of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0058] Example 1, referring to Figures 1-2As an embodiment of the present invention, a method for predicting long-term ice thickness is provided, comprising the following steps S1 to S5:

[0059] S1. Obtain the icing status dataset, preprocess the icing status dataset to obtain a standardized dataset;

[0060] S2. Construct a time-aware neural network model based on a standardized dataset;

[0061] S3. Based on time-varying and fixed data in a standardized dataset, a multi-channel fusion network is constructed using a time-aware neural network model to obtain a fusion prediction model.

[0062] S4. Train the fusion prediction model using the training strategy to obtain the trained prediction model;

[0063] S5. Use the trained prediction model to predict the icing thickness.

[0064] It should be noted that icing monitoring of transmission lines is a crucial aspect of power system disaster prevention, as icing is characterized by its long time span and wide coverage. In practical applications, icing thickness data retrieved via SAR is limited by satellite revisit cycles, resulting in irregular data acquisition intervals, ranging from as short as 1 day to as long as 12 days. This leads to significant temporal discontinuities in the icing thickness time series data. Furthermore, synchronizing multi-source information, such as meteorological data and line characteristic data, presents challenges in terms of time synchronization and dimensional consistency. Traditional LSTM networks assume equal-interval sampling, neglecting time interval information, which can lead to insufficient modeling of long-interval events and potential overemphasis on short-interval events. Moreover, existing purely data-driven models often exhibit physical paradoxes, such as predicting abnormally high ice thickness under no-precipitation conditions or continued increases in above-freezing temperatures. Therefore, developing an icing thickness prediction method that can handle irregularly timed data, integrate multi-source information, and adhere to physical laws is of significant practical importance.

[0065] Therefore, to address the aforementioned issues of discontinuous data time and lack of physical constraints, steps S1-S5 are employed to construct a preprocessing mechanism for multi-source data fusion, achieving time synchronization and standardized processing of asynchronously sampled data; a time-aware neural network model is built to accurately capture the temporal dynamic characteristics of icing thickness data through a time interval weight adjustment mechanism; a dual-channel fusion network architecture is established to effectively integrate dynamic meteorological information with static line characteristics; a phased training strategy and a physical constraint loss function are adopted to ensure that the prediction results conform to both data patterns and physical principles; and adversarial training is used to improve the model's robustness, achieving high-precision long-term series prediction and risk warning for icing thickness.

[0066] Example 2, refer to Figures 1-2As an embodiment of the present invention, based on the above embodiment, a long-term series prediction method for ice thickness is provided.

[0067] It is important to know that the icing status dataset includes historical icing thickness data, meteorological data, and line characteristic data. The meteorological data and historical icing thickness data are time-varying data, while the line characteristic data are fixed data.

[0068] In this embodiment of the application, step S1, the step of preprocessing the icing state dataset, includes:

[0069] The asynchronously sampled data is unified to a preset time grid using interpolation methods;

[0070] Specifically, the interpolation method used is cubic spline interpolation, which is expressed as follows:

[0071] f(t) = a i (tt i ) 3 +b i (tt i ) 2 +c i (tt i )+d i ;

[0072] In the formula, f(t) represents the interpolated data value; t i The time for the known data points; a i b i c i d i The coefficients are those of the cubic spline function; the coefficients are determined by ensuring continuity at the data points and continuity of the first and second derivatives.

[0073] It should be noted that, due to the limitations of satellite revisit cycles, SAR inversion data for icing thickness has irregular acquisition intervals, ranging from as short as 1 day to as long as 12 days. Meteorological data, on the other hand, is typically sampled daily or hourly, resulting in poor time synchronization. Cubic spline interpolation, by constructing piecewise cubic polynomials, can accurately fill in time intervals while maintaining data smoothness, unifying all data onto a time grid of Δt = 1 day, thus providing continuous time-series input for subsequent time-aware neural networks.

[0074] Standardization methods were used to standardize the sequence data.

[0075] Specifically, the standardization method adopted is the Z-score standardization method, which is expressed in the following form:

[0076]

[0077] In the formula, x* denoted as the standardized feature value; x is the original feature value; μ is the mean of the data in this dimension; and σ is the standard deviation of the data in this dimension.

[0078] For example, for historical data on temperature, humidity, wind speed, precipitation, air pressure, and icing thickness in meteorological data, calculate their mean values ​​μ respectively. temp and standard deviation σ temp :

[0079]

[0080] Where N is the number of data samples, T i Let be the i-th temperature observation value.

[0081] It should be noted that the ice thickness data is measured in millimeters, while meteorological data such as temperature, humidity, wind speed, precipitation, and air pressure have significant dimensional differences. Directly inputting these into the neural network can lead to numerical instability. Z-score standardization transforms the data in each dimension into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions and avoiding the dominant role of large numerical features in model training. It can also effectively handle the negative impact of outlier data on model performance.

[0082] In an alternative implementation, the interpolation method in step S1 can also be Lagrange interpolation, which fills in the data gaps by constructing Lagrange basis functions. This method is suitable for scenarios with few data points and requires precise passage through known points.

[0083] In another optional implementation, the standardization method in step S1 can also be the Min-Max standardization method, which scales the data to the [0,1] interval, and is suitable for situations where the data distribution is relatively uniform and the relative relationship of the original data needs to be maintained.

[0084] In this embodiment of the application, step S2, the construction steps of the time-aware neural network model, include A1 to A3:

[0085] A1. Adjust the short-term memory component by weighting the time intervals of adjacent data to obtain the reduced short-term memory.

[0086] A2. Combine long-term memory with reduced short-term memory to form a complete memory component;

[0087] A3. Construct a standard gating structure based on the complete memory components.

[0088] Specifically, in step A1, the time interval weight adjustment of the short-term memory component is expressed as follows:

[0089]

[0090] In the formula, For the short-term memory component at the current time step t; The short-term memory component of the previous time step; For reduced short-term memory after time interval weighting; W d b d These are the weight matrix and bias vector, respectively; Δt is the time interval between adjacent ice thickness data; g(·) is a non-increasing function, specifically in the form g(Δt)=1 / Δt.

[0091] It should be noted that when the time interval Δt is large (e.g., the satellite revisit cycle reaches 12 days), the value of g(Δt) is small, which reduces the contribution of short-term memory from the previous time step to the current prediction; when the time interval is small (e.g., 1-2 days), the value of g(Δt) is large, maintaining a strong time dependence. This design conforms to the physical laws of icing thickness variation, that is, the longer the time interval, the weaker the influence of preceding data.

[0092] Specifically, in step A2, the combination of long-term memory and reduced short-term memory is represented as follows:

[0093]

[0094]

[0095] In the formula, For long-term memory components; C t-1 This represents the complete memory state of the previous time step; This is the combined complete memory component, which serves as the input to the subsequent gating structure.

[0096] For example, for time series data on ice thickness, when Δt = 1 day, g(Δt) = 1, and short-term memory is fully preserved; when Δt = 12 days, g(Δt) = 1 / 21 ≈ 0.083, and short-term memory is significantly weakened, with prediction mainly relying on long-term memory.

[0097] Specifically, in step A3, the detailed mathematical expression of the LSTM gating structure is as follows:

[0098]

[0099] In the formula, f t i t o t , , represent the forget gate, input gate, and output gate, respectively; σ is the sigmoid activation function; W* and U* are the weights, and b* is the bias.

[0100] In an alternative implementation, the non-increasing function g(Δt) in step S2 is g(Δt).

[0101] g(Δt) can also be expressed in the form of exponential decay: g(Δt) = e -λΔt , where λ is the decay coefficient, which can more smoothly handle the impact of time interval changes on memory.

[0102] In another alternative implementation, the time-aware neural network model in step S2 can also adopt a gated recurrent unit (GRU) structure, which simplifies computational complexity by updating and resetting gates, making it suitable for application scenarios with limited computing resources.

[0103] In this embodiment of the application, step S2, the step of constructing the multi-channel fusion network, includes B1 to B3:

[0104] B1. Input the time-varying data into a time-series channel composed of a time-aware neural network for processing to obtain time-series features;

[0105] Specifically, the processing procedure for the timing channel is represented as follows:

[0106]

[0107] In the formula, h LSTM H is the timing feature vector output by the timing channel; t For historical data on icing thickness at time step t; X t The data represents meteorological data at time step t, including temperature, humidity, wind speed, precipitation, and air pressure; C represents timestamp information; T represents the time series length; and L represents the dimension of the LSTM hidden layer. This is the output of the i-th hidden unit at the final time step.

[0108] It should be noted that the time-series channel is specifically designed to process dynamic information that changes over time, using a time-aware LSTM to capture the temporal dependency between meteorological conditions and icing thickness. The input sequence has a dimension of D = 1 + 6 = 7 (1-dimensional historical icing thickness data + 6-dimensional meteorological data). After processing by the time-series channel, a fixed-dimensional feature representation is obtained, effectively compressing the time-series information while retaining key dynamic characteristics.

[0109] B2. Input fixed data into a static channel composed of a fully connected neural network for processing to obtain static features;

[0110] The specific form of the static channel processing is as follows:

[0111] h FC =ReLU(w·[diameter, material coefficient, altitude]+b);

[0112]

[0113] In the formula, hFC is the static feature vector output by the static channel; W is the weight matrix with dimension M×3; b is the bias vector; ReLU is the modified linear activation function; M is the number of hidden units in the fully connected layer; This is the output of the j-th hidden unit.

[0114] For example, for a certain transmission line with a conductor diameter of 25.2mm, a material coefficient of 0.85 (steel-cored aluminum stranded wire), and an altitude of 1200m, after standardization processing, the data is input into the static channel:

[0115] Input vector = [0.42, -0.35, 1.15];

[0116] h FC =ReLU(W 128×3 ·[0.42,-0.35,1.15]+b 128 );

[0117] Among them, W 128×3 This represents the weight matrix for outputting 128-dimensional static features.

[0118] B3. The temporal and static features are concatenated and fused, and the prediction results are generated through a fully connected layer;

[0119] The process of feature fusion and prediction result generation is represented as follows:

[0120] h fusion =[h LSTM ,h FC ];

[0121]

[0122] In the formula, h fusion The concatenated fused feature vector has a dimension of (L+M)×1; [·,·] represents the feature concatenation operation; haty is the final predicted icing thickness; W output The output layer weight matrix has a dimension of 1×(L+M); b output This is the output layer bias.

[0123] It should be noted that the feature concatenation operation linearly combines dynamic temporal features with static physical features in the feature space, forming a joint representation that takes into account both temporal evolution trends and physical constraints. The output layer maps the high-dimensional fused features to a one-dimensional predicted ice thickness value through a linear transformation, with weights W... output During training, the system automatically learns the optimal combination weights of temporal and static features.

[0124] In this embodiment of the application, step S4, the step of training the fusion prediction model, includes C1 to C3:

[0125] C1. The first stage involves training using a basic loss function until the fusion prediction model learns the basic data patterns. The basic loss function is the mean square error between the predicted value and the true value.

[0126] The base loss function uses mean squared error loss, and its form is as follows:

[0127]

[0128] In the formula, y represents the icing thickness predicted by the model. t The actual icing thickness value is represented by T, and the number of training samples is represented by T. The first phase of training continues until the validation set loss no longer decreases for 10 consecutive epochs, ensuring that the model fully learns the basic mapping relationship between icing thickness and meteorological factors and line characteristics.

[0129] It should be noted that the first phase of training enabled the model to capture the basic data patterns of ice thickness changes and establish a preliminary correlation between meteorological conditions and ice conditions. Through learning from a large amount of historical data, the model gradually mastered the influence patterns of factors such as temperature, humidity, and wind speed on ice thickness, laying the foundation for the subsequent introduction of physical constraints.

[0130] C2. In the second stage, a physical rule loss function is added to force the learning of physical laws. The physical rule loss function is used to punish predictions that violate physical laws.

[0131] Specifically, the physical rule loss function is constructed based on the physical constraints of icing growth, mainly including three aspects: temperature constraint, precipitation constraint, and rate of change constraint. When the ambient temperature is above 0℃, the icing thickness should not continue to increase; when there is no precipitation and no other snow and ice replenishment, the icing thickness should not surge abnormally; the rate of change of icing thickness should conform to the physical growth model specified in the IEC 60826 standard.

[0132] It should be noted that the introduction of the physical rule loss function can effectively avoid the physical paradoxes that occur in purely data-driven models. For example, when the model predicts that the ice thickness will still increase under sunny and hot weather, the physical constraint term will penalize such predictions that violate common sense, forcing the model to learn prediction patterns that conform to physical laws, and significantly improving the credibility and practicality of the prediction results.

[0133] C3. The third stage incorporates adversarial training, and finally outputs the trained prediction model.

[0134] The adversarial training process includes steps C3.1 to C3.4:

[0135] C3.1 Add Gaussian noise to the input data and construct data samples under extreme weather conditions;

[0136] C3.2 Input the data with Gaussian noise and the data samples under extreme weather conditions into the fusion prediction model for training;

[0137] C3.3 Update model parameters using the backpropagation algorithm;

[0138] C3.4. Set the intensity and frequency of adversarial training to improve the robustness of the fusion prediction model.

[0139] For example, in step C3.1, Gaussian noise is added by superimposing random noise with a mean of 0 and a standard deviation of 10% of the original data's standard deviation onto the original meteorological data, simulating sensor measurement errors and data transmission interference. Extreme weather conditions include severe meteorological environments such as abnormally low temperatures (below -30℃), blizzards (precipitation exceeding 50mm / day), and strong winds (wind speed exceeding 15m / s).

[0140] It should be noted that adversarial training, by deliberately introducing noise and extreme conditions, forces the model to learn stable predictive capabilities under uncertain environments. The proportion of noisy data is set at 20% of the total training data, and extreme weather samples account for 15%, ensuring that the model's robustness is improved without affecting the prediction accuracy of normal data. The training intensity is increased gradually, with a small noise amplitude initially, increasing progressively as training progresses, ultimately enabling the model to maintain reliable predictive performance in complex real-world environments.

[0141] In an alternative implementation, the physical rule loss function in step S4 can also be constructed based on the law of conservation of energy, thereby further enhancing the physical consistency of the model by constraining the energy balance relationship during the icing process.

[0142] In another alternative implementation, the adversarial training in step S4 can also adopt the idea of ​​generative adversarial networks, constructing generators and discriminators for adversarial learning, and improving the model's adaptability to abnormal situations through the game process.

[0143] In this embodiment of the application, step S5, which involves predicting the icing thickness using the trained prediction model, includes steps D1 to D4:

[0144] D1. Obtain real-time meteorological data and line characteristic parameters, and standardize the real-time data according to the same preprocessing method as the training phase.

[0145] The acquisition and preprocessing of real-time data includes collecting meteorological parameters such as temperature, humidity, wind speed, precipitation, and air pressure at the current moment from meteorological monitoring stations and line monitoring systems, as well as line characteristic parameters such as conductor diameter, material coefficient, and altitude.

[0146] It should be noted that standardization must use the statistical parameters saved during the training phase to ensure that real-time data and training data are within the same numerical range, avoiding prediction bias caused by differences in data distribution. Simultaneously, a data quality check mechanism should be established to identify and handle outliers, such as extreme readings caused by temperature sensor malfunctions or data loss due to communication interruptions.

[0147] D2. Input the preprocessed real-time meteorological data and historical icing thickness data into the time series channel and input the line characteristic parameters into the static channel. Perform forward propagation calculation through the trained multi-channel fusion network.

[0148] The forward propagation computation process of the multi-channel fusion network includes combining preprocessed real-time meteorological data with historical icing thickness data from the past 7 days to form a time series, which is then input into the time series channel for time-aware LSTM computation; inputting line characteristic parameters into the static channel for fully connected network processing; and finally splicing and fusing the output features of the two channels.

[0149] For example, in the forecasting process for a 110kV transmission line on January 15, 2025, the time-series channel input includes the meteorological data for that day [temperature -5℃, humidity 85%, wind speed 3.2m / s, precipitation 12mm, air pressure 1013hPa] and the historical data of icing thickness for the previous 7 days [2.1mm, 3.5mm, 4.2mm, 5.8mm, 7.1mm, 8.3mm, 9.6mm]; the static channel input includes the conductor diameter 25.2mm, material coefficient 0.85, and altitude 1200m.

[0150] D3. Obtain the predicted value of the icing thickness, perform denormalization on the prediction result to restore the original dimension, and conduct risk assessment and early warning judgment based on the preset safety threshold.

[0151] Risk assessment is based on the power industry's icing safety standards, with tiered early warning systems. When the predicted ice thickness is less than 10mm, it is considered a safe state, and normal monitoring is maintained. When the ice thickness is between 10-20mm, a yellow warning is issued, and the frequency of inspections is increased. When the ice thickness is between 20-30mm, an orange warning is issued, and de-icing equipment is activated. When the ice thickness exceeds 30mm, a red warning is issued, and emergency power outage measures are implemented.

[0152] D4. Output the prediction results in time series format.

[0153] For example, the output of a certain prediction is: the icing thickness will increase from the current 9.6 mm to 15.2 mm in the next 24 hours, with a growth rate of 0.23 mm / h, a prediction confidence level of 92%, a risk level of yellow warning, and recommended measures to increase the frequency of line inspections to once every 6 hours and prepare de-icing equipment.

[0154] It should be noted that the time series output also includes prediction uncertainty quantification. The prediction confidence interval is calculated using the Monte Carlo dropout method, and a manual review mechanism is automatically triggered when the prediction uncertainty exceeds a preset threshold. Simultaneously, a historical tracking and verification mechanism for prediction results is established to continuously evaluate the model's prediction accuracy and provide data support for model optimization and parameter tuning.

[0155] In an alternative implementation, the risk assessment in step S5 can also be combined with the importance of the line and the load level to set differentiated thresholds, and adopt stricter early warning standards for key transmission channels to ensure the safe and stable operation of the power grid.

[0156] In another optional implementation, the prediction results output in step S5 can also be integrated with visualization functions to generate an ice thickness change trend map and a risk distribution heat map, providing an intuitive decision support interface for operation and maintenance personnel.

[0157] In summary, the use of a time-aware LSTM network to address the temporal discontinuity of SAR inversion data, by introducing a time interval weighting mechanism, enables the model to accurately model the temporal dynamics of icing thickness data, significantly improving the prediction accuracy of long-term series compared to traditional LSTM networks. Secondly, the dual-channel fusion network architecture organically combines dynamic meteorological factors with static line characteristics. The time-series channel captures the dynamic correlation between meteorology and icing, while the static channel encodes the physical properties of the conductor; their collaborative operation greatly improves prediction accuracy. Simultaneously, the multi-source data fusion strategy fully utilizes the high-resolution SAR inversion results and rich meteorological auxiliary information, overcoming the limitations of a single data source.

[0158] This invention innovatively combines physical constraints with deep learning. Through a phased training strategy and a composite loss function design, it ensures that the prediction results conform to both data patterns and physical common sense. Adversarial training further enhances the model's robustness to noise and extreme weather, enabling the system to maintain stable and reliable prediction performance even in real-world complex environments. Overall, this invention provides a high-precision, high-reliability solution for predicting icing thickness in power systems, enabling accurate risk identification and early warning, significantly improving the safety of power grid operation, and reducing the risk of power outages and economic losses caused by icing.

[0159] Example 3 illustrates a schematic scheme for a long-term ice thickness prediction method. It should be noted that the technical solution of this long-term ice thickness prediction system belongs to the same concept as the technical solution of the aforementioned long-term ice thickness prediction method. Details not described in detail in this embodiment can be found in the description of the technical solution of the aforementioned long-term ice thickness prediction method.

[0160] This embodiment also provides a long-term series prediction system for icing thickness, including a data processing module, a model building module, a model training module, and a prediction module:

[0161] The data processing module is responsible for collecting the icing status dataset and performing preprocessing operations on the acquired icing status dataset.

[0162] The model building module designs and builds a time-aware neural network model based on the standardized dataset;

[0163] The model training module formulates and executes the training strategy to train the constructed fusion prediction model;

[0164] The prediction module predicts the icing thickness based on the input data.

[0165] This embodiment also provides an electronic device suitable for long-term ice thickness prediction, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the long-term ice thickness prediction method proposed in the above embodiment.

[0166] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for predicting long-term ice thickness as proposed in the above embodiments.

[0167] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for predicting long-term ice thickness proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0168] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0169] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting long-term ice thickness, characterized in that, Includes the following steps: Obtain the icing status dataset, and preprocess the icing status dataset to obtain a standardized dataset; Construct a time-aware neural network model based on the standardized dataset; Based on the time-varying and fixed data in the standardized dataset, a multi-channel fusion network is constructed using the time-aware neural network model to obtain a fusion prediction model. The fusion prediction model is trained using a training strategy to obtain a trained prediction model. The trained prediction model is used to predict the thickness of the ice layer.

2. The method for predicting long-term ice thickness as described in claim 1, characterized in that, The steps for preprocessing the icing state dataset include: The asynchronously sampled data is unified to a preset time grid using interpolation methods; Standardization methods were used to standardize the sequence data. The icing status dataset includes historical icing thickness data, meteorological data, and line characteristic data. The meteorological data and historical icing thickness data are time-varying data, while the line characteristic data are fixed data.

3. The method for predicting long-term ice thickness as described in claim 2, characterized in that, The steps for constructing the time-aware neural network model include: The short-term memory component is adjusted by weighting the time intervals between adjacent data to obtain a reduced short-term memory. The long-term memory is combined with the reduced short-term memory to form a complete memory component; A standard gating structure is constructed based on the complete memory components.

4. The method for predicting long-term ice thickness as described in claim 3, characterized in that, The steps for constructing the multi-channel fusion network include: Time-varying data is input into a time-series channel composed of a time-aware neural network for processing to obtain time-series features; Fixed data is input into a static channel composed of a fully connected neural network for processing to obtain static features; The temporal and static features are concatenated and fused, and a prediction result is generated through a fully connected layer.

5. The method for predicting long-term ice thickness as described in claim 4, characterized in that, The steps for training the fusion prediction model include: The first stage involves training using a basic loss function until the fusion prediction model learns the basic data patterns. The second stage incorporates a physical rule loss function to force the learning of physical laws; The third stage incorporates adversarial training, and finally outputs the trained prediction model. The basic loss function is the mean square error between the predicted value and the true value, and the physical rule loss function is used to penalize predictions that violate physical laws.

6. The method for predicting long-term ice thickness as described in claim 5, characterized in that, The steps of the adversarial training include: Add Gaussian noise to the input data and construct data samples under extreme weather conditions; Data with Gaussian noise and data samples under extreme weather conditions are input into the fusion prediction model for training; Update model parameters using the backpropagation algorithm; Adjust the intensity and frequency of adversarial training to improve the robustness of the fusion prediction model.

7. The method for predicting long-term ice thickness as described in claim 6, characterized in that, The steps for predicting icing thickness using a trained prediction model include: Acquire real-time meteorological data and line characteristic parameters, and standardize the real-time data using the same preprocessing method as in the training phase; The preprocessed real-time meteorological data and historical icing thickness data are input into the time series channel and the line characteristic parameters are input into the static channel. Forward propagation calculation is performed through the trained multi-channel fusion network. Obtain the predicted value of ice thickness, perform denormalization on the prediction result to restore the original dimension, and conduct risk assessment and early warning judgment based on the preset safety threshold. The prediction results will be output in time series format.

8. A long-term series prediction system for icing thickness, using the method described in any one of claims 1-7, characterized in that, It includes a data processing module, a model building module, a model training module, and a prediction module: The data processing module is responsible for collecting the icing status dataset and performing preprocessing operations on the acquired icing status dataset. The model building module designs and builds a time-aware neural network model based on the standardized dataset; The model training module formulates and executes a training strategy to train the constructed fusion prediction model; The prediction module predicts the icing thickness based on the input data.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the long-term ice thickness prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the long-term ice thickness prediction method according to any one of claims 1 to 7.

Images