Icing risk early warning method based on multi-model fusion and residual time sequence characteristic analysis

Abstract

The invention relates to the technical field of disaster prevention and reduction of a power system, and discloses an icing risk early warning method based on multi-model fusion and residual time sequence characteristic analysis, which comprises the following steps: collecting meteorological data of a line area in real time, removing abnormal values through secondary judgment of a Pauta criterion and a trend, and standardizing; adopting a TEROL algorithm to screen high-weight key features; running SWD-BP, MUL-GRNN and ELM models in parallel, constructing a dynamic weight by combining DSI, an independence weight method and an entropy weight method, and calculating a final meteorological predicted value; generating a prediction residual signal, extracting time domain features such as a mean value and a peak value, and constructing a residual feature matrix through a sliding window; and inputting an LSTM model to process a time sequence dependency relationship, and judging an icing risk level. According to the method, meteorological prediction is optimized through multi-model dynamic fusion, and deviation is analyzed and corrected in combination with residual time sequence characteristics, so that the problem of weak generalization ability of a single model is effectively solved, and the accuracy of icing risk early warning is obviously improved.

Description

Technical Field

[0001] This invention relates to the field of disaster prevention and mitigation technology in power systems, specifically to an icing risk early warning method based on multi-model fusion and residual time series feature analysis. Background Technology

[0002] Icing on transmission lines is one of the major natural disasters affecting the safe operation of power grids. In low-temperature, rainy, snowy, and freezing weather, the continuous accumulation of icing on lines can lead to conductor galloping, insulator flashover, and in severe cases, even tower collapse and line breakage, threatening the stability of the power system. Therefore, establishing a timely and accurate icing risk early warning mechanism is of great significance for guiding power grid operation and maintenance as well as disaster prevention and mitigation.

[0003] Current icing warning methods primarily rely on meteorological monitoring data combined with physical or statistical learning models for prediction. However, field-collected meteorological data typically contains numerous dimensions and is susceptible to noise from sensor malfunctions or environmental interference. Existing data processing methods are often coarse-grained in dealing with this data, failing to effectively retain true extreme weather characteristics while eliminating sensor erroneous data. Furthermore, unfiltered high-dimensional redundant features not only increase the computational burden on the model but also introduce irrelevant interference, affecting the input quality of subsequent prediction models.

[0004] In terms of predictive model construction, existing technologies mostly employ single neural networks or traditional physical mechanics models. Physical models involve complex fluid dynamics and thermodynamic calculations, which are difficult to meet the timeliness requirements of real-time dynamic early warning for power grids. On the other hand, single-structure statistical learning models often lack generalization ability and robustness when facing complex and variable micro-meteorological environments in mountainous areas. They are prone to getting trapped in local optima, resulting in significant fluctuations in the prediction accuracy of the model under specific meteorological conditions, making it difficult to guarantee the stability of the early warning results.

[0005] Furthermore, icing is an evolutionary process with a significant cumulative effect over time. Existing early warning strategies typically map numerical weather prediction data directly to icing thickness, ignoring the inherent biases in the weather forecast data itself. These methods fail to fully exploit the temporal dynamics inherent in the prediction errors (residuals) and cannot utilize the statistical distribution characteristics of model biases to correct risk assessment results in reverse. Consequently, when weather forecasts are biased, the accuracy of icing warnings is difficult to further improve. Summary of the Invention

[0006] To address the limitations of existing transmission line icing risk prediction methods, such as their simplistic model structure, insufficient generalization ability to extreme weather conditions, and failure to fully utilize statistical features in meteorological forecast errors, resulting in limited accuracy in icing warnings, this invention provides an icing risk warning method based on multi-model fusion and residual time-series feature analysis. This method improves the stability of meteorological parameter predictions through the fusion of multi-mechanism neural networks and extracts and classifies time-series features from the prediction residuals, thereby determining the level of icing risk.

[0007] To achieve the above objectives, this invention provides the following technical solution: an icing risk early warning method based on multi-model fusion and residual time-series feature analysis, comprising the following steps:

[0008] S1. Real-time collection of meteorological data for the target route area as actual observation values, identification and removal of outliers in the actual observation values, execution of standardization operations, and construction of an initial dataset; wherein, the meteorological data includes ambient temperature, daily minimum temperature, relative humidity, saturated humidity, absolute humidity, and wind speed;

[0009] S2. The TEROL algorithm is used to evaluate and calculate the meteorological features in the initial dataset, and a preset number of meteorological features with the highest weight values ​​are selected to construct a key meteorological feature dataset.

[0010] S3. Input the key meteorological feature dataset into the SWD-BP meteorological prediction model, MUL-GRNN meteorological prediction model and ELM meteorological prediction model for parallel processing, output the prediction results of each sub-model, and calculate the final meteorological prediction value based on the dynamic weight combination strategy.

[0011] S4. Calculate the difference between the final meteorological forecast value and the actual observed value to generate a residual signal, extract the time-domain features of the residual signal, process the time-domain features using a sliding window algorithm, and construct a residual feature matrix that reflects the dynamic change law of the residual; wherein, the time-domain features include mean, peak value, peak-to-peak value, sum of squares, and root mean square amplitude;

[0012] S5. Input the residual feature matrix into the pre-trained LSTM sequence classification model to capture the temporal dependencies in the residual feature matrix, and output the current icing risk level and warning information.

[0013] Preferably, in step S1, the specific process of identifying and removing outliers from the actual observations and performing standardization includes: initially screening the actual observations based on the Laida criterion, calculating the mean and standard deviation, and marking data exceeding a set threshold as suspected outliers; performing secondary discrimination on the suspected outliers, calculating the rate of change of data before and after the time of the suspected outlier's occurrence, and comparing it with the distribution of meteorological extreme values ​​in the same historical period; if it conforms to the meteorological physical evolution law and is within the tolerance range of historical extreme values, the suspected outlier is retained as a true extreme value; otherwise, it is determined as a mismeasured value; removing the mismeasured values ​​from the data sequence, and normalizing the remaining data after removal to generate the initial dataset.

[0014] Preferably, in step S2, step S2 specifically includes: randomly selecting samples from the initial dataset, finding correctly guessed nearest neighbors among samples of the same type, and finding incorrectly guessed nearest neighbors among samples of different types; calculating the normalized distance between the sample and the correctly guessed nearest neighbor, and between the sample and the incorrectly guessed nearest neighbor, on each meteorological feature dimension; iteratively updating the weight values ​​of each meteorological feature using the normalized distance until a preset maximum number of iterations is reached to obtain the final weight of each feature; sorting the final weights, selecting a preset number of features with the highest weight values ​​as key features, and using the key features to construct the key meteorological feature dataset.

[0015] Preferably, in step S3, the process of constructing the SWD-BP meteorological prediction model and the MUL-GRNN meteorological prediction model includes: establishing a three-layer BP neural network, encoding the initial connection weights and thresholds of the BP neural network into position vectors, using the SWD algorithm to perform global iterative optimization on the position vectors, assigning the decoded optimal parameters to the BP neural network, and constructing the SWD-BP meteorological prediction model. A generalized regression neural network (GRNN) is established, defining a variable matrix to represent the state of the black ant colony, using the MUL algorithm to iteratively optimize the smoothing parameter (SPREAD) of the Gaussian function in the model layer of the generalized regression neural network (GRNN), and substituting the determined optimal SPREAD parameters into the generalized regression neural network (GRNN) to construct the MUL-GRNN meteorological prediction model.

[0016] Furthermore, the parameter values ​​for the next iteration are calculated using the SWD parameter optimization formula. In the MUL algorithm, the variable values ​​in the variable matrix are initially updated using the roulette wheel update formula, the material exchange update between black ant colonies is updated using the tunnel exchange update formula, and the tunnel existence probability and fulfillment distance rate are calculated using the parameter adjustment formula to balance the global and local search capabilities.

[0017] Preferably, in step S3, the specific steps for calculating the final meteorological forecast value based on the dynamic weight combination strategy include: obtaining the inverse normalized forecast results of each sub-model and dividing the forecast results into an initial group and a test group according to the time series; based on the test group data, statistically analyzing the distribution frequency of each sub-model's forecast value above and below a set threshold, and calculating the DSI weight based on the forecast frequency; performing error correlation analysis using the previous day's meteorological data and the initial group data, and calculating the initial objective weight based on the independence weight method; for the test group data, evaluating the information entropy value of each sub-model within the future forecast period, and calculating the objective weight based on the entropy weight method; combining the DSI weight based on the forecast frequency, the initial objective weight based on the independence weight method, and the objective weight based on the entropy weight method to construct a combined weight matrix, and using the combined weight matrix to perform a weighted summation of the forecast results of each sub-model to output the final meteorological forecast value.

[0018] Furthermore, in the process of calculating the weights, the DSI weight formula is used to calculate the DSI weights based on the prediction frequency, the negative correlation coefficient formula and the initial objective weight formula are used to calculate the initial objective weights based on the independence weight method, the entropy weight formula is used to calculate the objective weights based on the entropy weight method, and the combined weight formula is used to calculate the combined weight matrix.

[0019] Preferably, in step S4, the time-domain features of the residual signal are extracted using the following formula:

[0020] The arithmetic mean of the residual signal is calculated using the mean formula;

[0021] The maximum value of the residual signal is calculated using the peak value formula;

[0022] The difference between the maximum and minimum values ​​of the residual signal is calculated using the peak-to-peak value formula;

[0023] The sum of squares formula is used to calculate the total energy of the residual signal;

[0024] The root mean square amplitude of the residual signal is calculated using the root mean square amplitude formula.

[0025] Preferably, in step S5, the training process of the LSTM sequence classification model includes: constructing a loss function containing a regularization term, which is composed of the cross-entropy error function and the L2 regularization term, and taking half of the product of the transpose of the weight vector and the weight vector as the L2 regularization term; iteratively updating the learnable parameters of the LSTM sequence classification model using an improved stochastic momentum gradient descent strategy, and calculating the parameters for the next iteration using the stochastic momentum parameter update formula, wherein the stochastic momentum parameter update formula introduces a random perturbation factor into the momentum term.

[0026] Preferably, in step S5, the output icing risk level is based on the equivalent icing thickness. Divided into four levels:

[0027] Level 1 risk corresponds to no icing or minimal icing. ;

[0028] Level 2 risk, corresponding to light icing. ;

[0029] Level 3 risk, corresponding to moderate icing. ;

[0030] Level 4 risk corresponds to severe icing. .

[0031] This invention provides a method for early warning of icing risk based on multi-model fusion and residual time-series feature analysis. It has the following beneficial effects:

[0032] 1. This invention constructs a differentiated prediction model group including SWD-BP, MUL-GRNN and ELM, and adopts a dynamic weight combination strategy based on DSI, independence weight method and entropy weight method. This strategy can dynamically adjust the weight allocation according to the actual performance of each sub-model within different forecast lead times, avoiding the problem that a single model is prone to getting into local optima or insufficient generalization ability under specific meteorological conditions, thereby improving the accuracy and stability of meteorological parameter prediction.

[0033] 2. This invention utilizes the TEROL algorithm to perform feature filtering on multidimensional meteorological data. By calculating the normalized distance between the sample and its nearest neighbor and iteratively updating the feature weights, key features such as wind speed, temperature, and humidity are filtered out. Combined with an outlier handling mechanism based on the Laida criterion and secondary discrimination, redundant features and sensor mismeasurement data are effectively eliminated while retaining the real extreme meteorological values ​​that are crucial for icing formation, thereby improving the computational efficiency of subsequent models and the quality of input data.

[0034] 3. This invention introduces a residual time series feature analysis mechanism, which transforms the deviation between meteorological forecast values ​​and actual values ​​into a time-domain feature matrix containing information such as mean, peak value and energy. The LSTM network is used to process the time series correlation of the residual sequence. This method can use the statistical distribution law in the meteorological forecast deviation to help correct the risk judgment result, which solves the problem that the traditional method only relies on the meteorological forecast value and ignores the implicit information of the forecast error, and further improves the accuracy of icing risk warning. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a high-precision dynamic early warning method for icing risk based on multi-model fusion and residual time-series feature analysis according to an embodiment of the present invention.

[0036] Figure 2This is a schematic diagram showing the numerical distribution of the weights of various meteorological elements according to an embodiment of the present invention;

[0037] Figure 3 A comparison chart of MAPE error and MAX error for some models provided in embodiments of the present invention;

[0038] Figure 4 A comparison chart of RESM errors for some models provided in embodiments of the present invention. Detailed Implementation

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] See attached document Figure 1 This invention provides a high-precision dynamic early warning method for icing risk based on multi-model fusion and residual time-series feature analysis. This method relies on a complete early warning processing architecture. Logically, this architecture includes a meteorological data acquisition module, a data preprocessing module, a feature selection module, a fusion prediction module, a residual analysis module, and a risk early warning module. The meteorological data acquisition module acquires field environmental data; the data preprocessing module cleans and standardizes the data; the feature selection module reduces data dimensionality; the fusion prediction module integrates multiple neural network models to output meteorological forecast values; the residual analysis module extracts time-series features from the prediction error; and the risk early warning module determines the final risk level.

[0041] The method specifically includes the following steps:

[0042] S1. The meteorological data acquisition module collects various meteorological data of the target route area in real time, including ambient temperature, daily minimum temperature, relative humidity, saturated humidity, absolute humidity, and wind speed.

[0043] S2. The data preprocessing module receives various types of meteorological data, identifies outliers in the data according to the Laida criterion, and performs secondary discrimination of outliers by combining real-time data trends and historical data for the same period, eliminating false positives and retaining true extreme values. Subsequently, the processed data is standardized to construct an initial dataset. .

[0044] S3. The feature selection module uses the TEROL algorithm to filter the initial dataset. The features in the dataset are evaluated and calculated, and a predetermined number of key features with the highest weight values ​​are selected to construct a key meteorological feature dataset. .

[0045] S4. The fusion prediction module runs the SWD-BP meteorological prediction model, the MUL-GRNN meteorological prediction model, and the ELM meteorological prediction model in parallel, respectively, on the key meteorological feature datasets. The system processes the data, outputs the prediction results of each sub-model, and calculates the final meteorological forecast based on a dynamic weight combination strategy.

[0046] S5. The residual analysis module calculates the difference between the final meteorological forecast and the actual observed value, generates a residual signal, and extracts five time-domain features of the residual signal: mean, peak value, peak-to-peak value, sum of squares, and root mean square amplitude.

[0047] S6. The residual analysis module uses the sliding window algorithm to process five time-domain features and constructs a residual feature matrix that reflects the dynamic changes of the residuals. .

[0048] S7, the risk warning module will use the residual feature matrix The input is fed into a pre-trained LSTM sequence classification model, which captures the temporal dependencies in the feature matrix.

[0049] S8. The risk warning module determines the current icing risk level based on the output of the LSTM sequence classification model. The icing risk level includes no icing, light icing, moderate icing, and heavy icing, and outputs warning information through the interactive interface.

[0050] To enable those skilled in the art to better understand the technical solution of the present invention, the above steps will be described in detail below in conjunction with specific algorithm principles and implementation details.

[0051] Based on the overall architecture description in Part 1 above, this embodiment elaborates on the preprocessing of meteorological data and the key feature screening process based on the TEROL algorithm as follows.

[0052] In the data preprocessing stage, to ensure the data quality of the input model, this embodiment of the invention employs a dual screening mechanism combining statistical criteria and time-series trend discrimination. Specifically, the initial dataset is constructed through the following steps:

[0053] S201. Obtain raw meteorological data sequences. Various meteorological data for the target route area are collected in real time using meteorological monitoring devices. These meteorological data specifically include ambient temperature. Daily minimum temperature relative humidity saturated humidity absolute humidity and wind speed .

[0054] S202. Initial screening of outliers for various meteorological data based on the Raida criterion (3σ criterion). This step assumes that meteorological monitoring errors follow a normal distribution and calculates the mean μ and standard deviation σ of various meteorological data. If the monitoring data at a certain time... satisfy If a value is found to be an outlier, it will be marked as a suspected outlier. The specific calculation process of the Raida criterion is well-known in this field and will not be elaborated here.

[0055] S203. Secondary discrimination is performed on the initially screened suspected outliers. Both genuine extreme weather events and false readings caused by sensor malfunctions appear as outliers numerically. To preserve the true characteristics of extreme weather, this step combines real-time data trends with historical data from the same period. Specifically, the rate of change of data before and after the suspected outlier's occurrence is calculated, and the distribution of meteorological extreme values ​​for the same period in history (e.g., the same time period over the past 5 years) is retrieved. If the trend of the suspected outlier conforms to the laws of meteorological physical evolution, and the value is within a certain tolerance range of historical extreme values, it is determined to be a genuine extreme value and retained; conversely, if it exhibits isolated, impulsive abrupt changes without historical data support, it is determined to be a false reading and discarded.

[0056] S204. Construct the initial dataset. Normalize the various meteorological data after removing erroneous values ​​to eliminate the influence of units, thus constructing the initial dataset. Initial dataset It can be represented as .

[0057] Obtaining the initial dataset Subsequently, to improve the computational efficiency and accuracy of the subsequent prediction model, this embodiment of the invention employs the TEROL (Two-stage Error Rate based Overlap Level) algorithm to evaluate and filter data features. The TEROL algorithm iteratively updates feature weights, measuring the discriminative power of features for sample classification, thereby filtering out the key features containing the most information. This process specifically includes the following steps:

[0058] S301, Sample Initialization. Starting from the initial dataset... A sample is randomly selected from the middle. And search in similar samples The set of 10 nearest neighbors is denoted as _{nearestHit}. Simultaneously, search for outlier samples. The set of 1 few nearest neighbors that were guessed incorrectly is denoted as _{nearestmiss}. .

[0059] S302. Calculate the feature distance. For any feature in the initial dataset... Calculate samples Its nearest neighbor samples Normalized distance on this feature. Using... The formula calculates the normalized distance. The formula is:

[0060] ;

[0061] In the formula, Indicates sample And guess the nearest neighbor sample In features Normalized distance on; Indicates sample In features The specific value on; Indicates the first A sample of the nearest neighbor that was guessed. In features The specific value on; Indicates that both have characteristics The absolute difference on; Representation of features The maximum value in the entire dataset; Representation of features The minimum value in the entire dataset; Representation of features The range of values ​​(i.e., the range). Sample Mis guessed nearest neighbor samples In features Normalized distance on The same logical calculations are used.

[0062] S303, Feature Weight Update. The feature weight update formula is used to update the features. The weights are iteratively updated, and the feature weight update formula is:

[0063] ;

[0064] In the formula, Indicates the updated features The weight value; Indicates the features in the current iteration step The weight value; Indicates the number of nearest neighbor samples; Indicates the number of samples or the smoothing factor; This represents a weighted penalty term based on the nearest neighbor distance of the same class. The negative sign before this term indicates that the sample... The smaller the distance to its nearest neighbor of the same type, the lighter the penalty for the current weight; Indicates the sample category; Indicates sample The category to which it belongs; Indicate category Prior probabilities in the dataset; This is the category balance coefficient; The overall representation is a weighted gain term based on the separation degree of outlier samples; the greater the distance between outliers, the greater the weighted gain.

[0065] S304. Filter key features. Repeat steps S301 to S303 until the preset maximum number of iterations is reached. Iterate through the final weight values ​​of all features. The features were then sorted, and the four features with the highest weight values ​​were selected as key meteorological features.

[0066] See attached document Figure 2 As shown in the figure, after evaluation using the TEROL algorithm, the weight values ​​of each meteorological element show significant differences, with the horizontal axis values ​​ranging from 0 to 0.25. Specifically:

[0067] The bar corresponding to wind speed is the longest, with a weight value close to 0.25, indicating that it contains the most information.

[0068] The length of the bar for ambient temperature is slightly shorter than that for wind speed, and the value is also close to 0.25;

[0069] The weighting values ​​for relative humidity are between 0.20 and 0.25; the weighting value for daily minimum temperature is close to 0.20. This is followed by saturation humidity (approximately 0.20), absolute humidity, and specific humidity (approximately 0.15).

[0070] Factors such as solar radiation, sunshine duration, and dew point temperature have relatively low weights, and the weight of precipitation is even less than 0.05.

[0071] Based on the above appendix Figure 2 The weighted ranking results shown in this embodiment ultimately identify the key meteorological feature as: wind speed. Ambient temperature relative humidity Lowest temperature of the day .

[0072] S305. Construct a key meteorological feature dataset. Based on the screening results, extract the initial dataset. The corresponding columns are used to construct a key meteorological feature dataset. ,Right now This dataset This will serve as the direct input for the subsequent multi-model fusion prediction module. Through the TEROL algorithm's filtering process, the dimensionality of the model input is significantly reduced while retaining core information.

[0073] After the meteorological data preprocessing and key feature screening are completed, this embodiment constructs a fusion prediction architecture containing three different neural network mechanisms to process the key meteorological feature dataset in parallel. To overcome the limitations of single models in terms of nonlinear fitting ability and generalization performance, parallel prediction is carried out using the SWD-BP neural network model, the MUL-GRNN meteorological prediction model, and the ELM extreme learning machine model.

[0074] A SWD-BP meteorological forecasting model optimized based on the SWD algorithm is constructed. This model uses a BP (BackPropagation) neural network as the basic predictor, with a three-layer topology: the input layer has 4 nodes, corresponding to the four key meteorological features; the hidden layer has 5 nodes, used to extract the nonlinear mapping relationships between features; and the output layer has 1 node, corresponding to the output meteorological forecast value. Given that traditional BP neural networks are prone to getting trapped in local optima and are sensitive to initial weights, this embodiment introduces the SWD (Small World Discovery) algorithm to perform global iterative optimization of the initial weights and thresholds of the BP neural network. All initial connection weights and thresholds of the BP neural network are encoded as position vectors. The global optimum is found by iteratively updating the parameter positions. A maximum number of iterations is set during the parameter optimization process. The value is 1000, and the current iteration number is denoted as . The parameter values ​​for the next iteration are calculated using the SWD parameter optimization formula, which is:

[0075] ;

[0076] In the formula, Indicates the parameter value for the next iteration; This indicates the parameter value at the current iteration. Indicates the current iteration number; This indicates the maximum number of iterations set. for Random numbers between; when random numbers At that time, a nonlinear decay strategy based on iterative progress is used to update the parameters to maintain a large search step size in the early stages of the search; when At this point, a linear random perturbation strategy is employed to enhance local search capabilities. After the iteration concludes, the optimal parameters are... The data is decoded and assigned to a backpropagation (BP) neural network to predict the key meteorological features of the input, and the first meteorological forecast value is output. .

[0077] A MUL-GRNN weather forecasting model optimized based on the MUL algorithm is constructed. This model employs a generalized regressive neural network (GRNN), whose network structure includes an input layer, a model layer, a summation layer, and an output layer. The input layer has 4 neurons, and the number of neurons in the model layer is the same as the number of training samples. The model layer of the GRNN uses a Gaussian function as its transfer function, and its predictive performance is highly dependent on the smoothing parameter (SPREAD) of the Gaussian function. Therefore, this embodiment uses the MUL (Multi-Universal Parallel Linear) algorithm to iteratively optimize the SPREAD parameter.

[0078] The MUL algorithm updates the solution space by simulating the process of matter exchange and expansion in parallel universes. Define the variable matrix. , used to represent the state of the black ant colony (i.e., the candidate solution set), variable matrix for:

[0079] ;

[0080] In the formula, Indicates the number of variables; Indicates the number of black ant colonies; Indicates the first The first individual of the black ant colony There are several variable values. During the iteration process, each individual variable is first updated according to the roulette wheel update mechanism. The updated variable values ​​are calculated using the roulette wheel update formula, which is:

[0081] ;

[0082] In the formula, The first one to be updated The first black ant colony One variable; The first one selected by roulette The first black ant colony One variable; For the first The normalized inflation rate of each variable; It is a random value between [0,1].

[0083] Subsequently, the black ant colony exchanges matter through tunneling mechanisms to further update its bloat rate. The variable values ​​are calculated using a tunneling exchange renewal formula, which is:

[0084] ;

[0085] In the formula, Denotes the i-th iteration of the optimal black ant colony in the current iteration. Each parameter value; Indicates the probability that the tunnel exists; Indicates the fulfillment distance rate; , , All are random values ​​between [0,1]; and They represent the first The upper and lower boundaries of the search space for each parameter. and This is a key parameter in the MUL algorithm for balancing global and local search, and it changes dynamically with the number of iterations. The parameter adjustment formula is used to calculate... and The parameter adjustment formula is:

[0086] ;

[0087] ;

[0088] In the formula, and These represent the preset initial minimum and maximum values ​​of the probability of the tunnel existing, respectively; This represents the total number of iterations. This represents the current iteration number; for The adjustment parameters are then determined using the MUL algorithm. After identifying the optimal SPREAD parameters, they are fed into a generalized regressive neural network (GRNN) to predict key meteorological features, outputting a second meteorological forecast value. .

[0089] An ELM weather forecasting model was constructed. Extreme Learning Machine (ELM) was used as the third-path prediction model, with 4 nodes in the input layer, 20 nodes in the hidden layer, and 1 node in the output layer. During model initialization, the connection weight matrix between the input and hidden layers, as well as the bias values ​​of the hidden layer neurons, were randomly generated and remained unchanged during subsequent training. The sigmoid function was chosen as the activation function for the hidden layer neurons, and the output weights from the hidden layer to the output layer were calculated using the least squares method. The constructed ELM model was then used to predict key meteorological features, outputting the third-path weather forecast values. .

[0090] The meteorological forecast values ​​output by the above three sub-models , , The dataset is denormalized to restore it to a meteorological forecast dataset with actual physical dimensions. , , This data serves as the foundation for the subsequent dynamic combination weight calculation module.

[0091] After obtaining the prediction results output by the three meteorological prediction sub-models—SWD-BP meteorological prediction model, MUL-GRNN meteorological prediction model, and ELM meteorological prediction model—this embodiment adopts a fusion strategy based on three-dimensional dynamic weights to obtain high-precision final meteorological prediction values ​​and generate residual signals.

[0092] The weather forecast dataset is time-series grouped. The weather forecast datasets after inverse normalization of the three sub-models mentioned above are then obtained. , , The dataset was divided into an initial group and a test group. The initial group consisted of the first eight time periods in the time series, used to calculate the correlation with historical data; the remaining time periods formed the test group. It is used for dynamic weight adjustment based on future trends.

[0093] Calculate the DSI weight based on the prediction frequency. Combined with test group data, statistically analyze the first DSI weight within the 48-hour forecast lead time. In each time period, the number of methods in which the predicted value of each sub-model is above the set threshold is denoted as . The number of methods whose predicted values ​​are at or below a set threshold is denoted as . The weights are determined based on the frequency of the positive and negative distribution of the predicted values, and the DSI weight formula is used to calculate the DSI weights for the next 48 hours of forecast time. The DSI weight formula is:

[0094] ;

[0095] In the formula, For the first The sub-model in the first DSI weights for each time period; This refers to the number of methods in all sub-models at that time whose meteorological forecast values ​​are above a set threshold. This refers to the number of methods in which the meteorological forecast value in all sub-models at or below a set threshold for that time period. For the first Time sequence Weather forecasts from individual models.

[0096] Calculate the initial objective weights based on the independence weighting method. Compare the previous day's actual weather data with the initial group forecast data to assess the correlation between the prediction errors of each sub-model. The initial objective weights are calculated using the independence weighting method. First, calculate the negative correlation coefficients between each sub-model. The formula for calculating the negative correlation coefficient is revised as follows:

[0097] ;

[0098] In the formula, For the first Time sequence Weather forecasts from individual sub-models; For the first Time sequence The average weather forecast of each sub-model; For the first Time sequence Standardized meteorological forecasts from individual sub-models; This represents the average of the corresponding standardized values; The number of methods involved in the calculation (3 in this case).

[0099] Based on negative correlation coefficient The weight values ​​are calculated using the initial objective weight formula, which is:

[0100] ;

[0101] In the formula, For the first Initial objective weights for each sub-model; The number of sub-models is given. Objective weights based on the entropy weighting method are calculated. For the test group data, the amount of information provided by each sub-model within the next 48-hour forecast lead time is evaluated. Objective weights are calculated using the entropy weighting method. First, the meteorological data is standardized to obtain... The standardized formula is:

[0102] ;

[0103] Next, calculate the first... Time sequence The proportion of predictions from individual sub-models The formula for calculating specific gravity is:

[0104] ;

[0105] Finally, the weight values ​​are calculated using the entropy weight method formula, which is as follows:

[0106] ;

[0107] In the formula: The objective weight value is calculated using the entropy weight method; The length of the time series being evaluated; The number of sub-models. When At that time, it was stipulated .

[0108] Construct the combined weight matrix and generate the final predicted values. Taking into account prediction accuracy, model independence, and information entropy, the final combined weight matrix is ​​calculated using the combined weight formula. The formula for the combined weights is:

[0109] ;

[0110] In the formula, For the first Time sequence The final combined weights of the sub-models. Independence weights are calculated using the previous day's weather data. The DSI weights are calculated using forecast data for the next 48 hours. And entropy weight This enables 24-hour periodic dynamic updates of the combined weights.

[0111] Based on the combined weight matrix, the final meteorological forecast result is calculated using a weighted summation formula. The weighted summation formula is:

[0112] ;

[0113] In the formula, The final meteorological values ​​after combined weighting; These are the corresponding combined weights; For the test group The prediction data of each sub-model.

[0114] Generate residual signals. Then, use the final weather forecast results... Compared with actual observed meteorological data Perform differential calculations to obtain the difference between the two, i.e., the residual signal. When transmission lines experience icing, the predicted deviations of meteorological parameters exhibit specific nonlinear variations, and this residual signal... This forms the basis for subsequent time series feature analysis.

[0115] The residual signal between the meteorological forecast and the actual data was obtained. Subsequently, the residual signal contained nonlinear biases that the model failed to fit, and these biases were closely related to the microscopic changes in the meteorological and physical environment caused by icing of transmission lines.

[0116] This embodiment performs time-domain feature extraction and matrix construction on the residual signal to uncover its evolutionary patterns.

[0117] S501. Extract the time-domain features of the residual signal. To comprehensively characterize the fluctuation characteristics and energy distribution of the residual signal at different time scales, the mean, peak value, peak-to-peak value, sum of squares, and root mean square amplitude are selected as five key time-domain feature indicators. The currently extracted residual signal segment is defined as... ,in Let be the length of the signal segment. To account for the central tendency of the residual signal, a mean formula is used for calculation. The mean formula is:

[0118] ;

[0119] In the formula, Represents the arithmetic mean of the residual signal segments; Indicates the first segment in the fragment The residual values ​​of each sampling point; This represents the total number of sampling points within the segment. This indicator reflects the DC component of the prediction bias, i.e., a systematic trend of being too high or too low. For extreme fluctuations in the residual signal, a peak value formula is used for calculation. The peak value formula is:

[0120] ;

[0121] In the formula, This represents the maximum value within the residual signal segment. This index is used to capture the instantaneous maximum deviation caused by sudden meteorological factors (such as gusts of wind or sudden cooling). The peak-to-peak value formula is used to calculate the fluctuation range of the residual signal. The peak-to-peak value formula is as follows:

[0122] ;

[0123] In the formula, This represents the difference between the maximum and minimum values ​​in a residual signal segment. This index characterizes the oscillation amplitude of the residual signal within the statistical period and is highly sensitive to instability disturbances in meteorological parameters during the initial stage of icing. The energy intensity of the residual signal is calculated using the sum-of-squares formula, which is:

[0124] ;

[0125] In the formula, This represents the total energy of the residual signal segments. This index reflects the overall strength of the prediction bias; when increased icing on the line leads to abnormal changes in meteorological parameters, the energy of the prediction residuals usually increases significantly. To assess the waveform stability of the residual signal, the root mean square (RMS) amplitude formula is used for calculation. The RMS amplitude formula is as follows:

[0126] ;

[0127] In the formula, This represents the root mean square amplitude of the residual signal. By taking the square root of the signal amplitude, this indicator is more sensitive to small amplitude fluctuations than the conventional root mean square value, helping to identify weak signal characteristics in the early stages of icing.

[0128] S502. Construct a residual feature matrix based on a sliding window. To preserve the time dependence of the residual signal and adapt to the input requirements of the LSTM network, a sliding window technique is used to generate a continuous sequence of feature vectors.

[0129] Set the width of the sliding window to (For example =8, meaning it covers data from the past 8 time points), with a sliding step size of 1. For a residual signal sequence of total length n, a time window is extracted starting from the m-th time point. The residual data within is used as the current segment .

[0130] The current segment is calculated using the five formulas in step S501. mean Peak Peak-to-peak value Sum of squares E and root mean square magnitude These five eigenvalues ​​are combined to form the eigenvector at time t. The feature vector generation function is expressed as:

[0131] ;

[0132] In the formula: for The feature vector at time step 1 has a dimension of 1. ; Indicates from arrive The residual subsequence at time step; This represents the set mapping of the five time-domain feature extraction algorithms mentioned above.

[0133] In chronological order, the feature vectors calculated at all times... Stack them to construct the final residual feature matrix. Feature matrix The structure is represented as:

[0134] ;

[0135] In the formula: for A matrix of dimensionality; These correspond to the feature vectors at different times. This feature matrix... The residual signal is transformed from a single one-dimensional numerical sequence into a time series matrix containing multi-dimensional statistical properties, which serves as the direct input to the subsequent LSTM sequence classification model.

[0136] In obtaining residual feature matrices containing rich dynamic information Subsequently, this example uses deep learning technology to classify and discriminate sequence features in order to achieve accurate early warning of the risk of icing on transmission lines.

[0137] An LSTM sequence classification model is constructed. Given that the residual signal is long-period time-series data and the icing process has a significant cumulative effect, this embodiment constructs a Long Short-Term Memory (LSTM) network to capture long-term dependencies in the feature matrix. The network architecture of this LSTM sequence classification model includes an input layer, an LSTM layer, a fully connected layer, and a softmax classification layer. The input layer is configured with a receive dimension of... residual characteristic matrix ,in The time step size of the sliding window is 5, corresponding to the five temporal feature dimensions: mean, peak value, peak-to-peak value, sum of squares, and root mean square magnitude. The LSTM layer, as the core processing unit, regulates the information flow through its internal forget gate, input gate, and output gate mechanisms. The time steps are set to 6000 to cover a long temporal evolution cycle sufficient to characterize the icing formation process. The fully connected layer maps the high-dimensional hidden state vector output by the LSTM layer to the sample label space. The Softmax classification layer, as the output, outputs the probability distribution of the current state belonging to each icing risk level.

[0138] A loss function incorporating a regularization term is constructed. To improve the model's generalization ability and prevent overfitting on the training data, this embodiment introduces a regularization term on top of the traditional cross-entropy loss function. Regularization term. The overall objective function value is calculated using the regularization loss function formula, which is:

[0139] ;

[0140] Among them, the regularization term The calculation formula is:

[0141] ;

[0142] In the formula, This represents the total error function after introducing regularization; This represents the set of all learnable parameters of the model (including weights and biases). Represents the basic cross-entropy error function; The regularization coefficient is used. This represents the weight vector in the model; This represents the transpose of the weight vector.

[0143] The stochastic momentum gradient descent algorithm is used to optimize the model parameters. Traditional gradient descent algorithms are prone to getting trapped in local minima. Therefore, this embodiment proposes an improved stochastic momentum gradient descent strategy to update the model parameters. This strategy introduces a random perturbation factor into the momentum term. The parameters are iteratively updated using a random momentum parameter update formula, which is:

[0144] ;

[0145] In the formula, Indicates the first Model parameters after the next iteration; Indicates the first Model parameters at the next iteration; Indicates the first Model parameters at the next iteration; The learning rate; This indicates the target loss function with the current parameters. The gradient vector at that point; This is the random momentum coefficient, which takes the value of a random number between 0 and 1. By introducing a random coefficient... By utilizing historically updated inertial information and applying random perturbations, the model's ability to explore the parameter space is increased, helping it escape local extrema.

[0146] The system determines the icing risk level and outputs early warning information. After processing the input residual feature matrix, the trained LSTM sequence classification model outputs the probability that the current transmission line status belongs to each risk level. The system determines the current icing risk level based on the maximum probability principle. In this embodiment, four risk level labels are pre-defined based on equivalent icing thickness for supervised model training and interpretation of early warning meanings. The specific classification criteria are as follows:

[0147] Level 1 Risk (No or Minimal Icing): Corresponding Equivalent Icing Thickness ;

[0148] Level 2 risk (light icing): corresponding equivalent icing thickness ;

[0149] Level 3 risk (moderate icing): corresponding equivalent icing thickness ;

[0150] Level 4 Risk (Severe Icing): Corresponding equivalent icing thickness Early Warning Response and Interaction. The risk early warning module displays the judgment results through a visual interactive interface. When the judgment result is a level three or four risk, the system highlights the alarm area on the interface and pushes the early warning information (including the specific tower location, the corresponding risk level, and the icing thickness range) to the maintenance personnel's terminal through the communication interface to support emergency decision-making.

[0151] To verify the technical effectiveness of the icing early warning method based on multi-model dynamic fusion and residual time series analysis proposed in this invention, this embodiment selects real meteorological monitoring data of a transmission line in a certain region during historical icing cycles as experimental samples, with a sampling frequency of once per hour. The fusion prediction model proposed in this invention is compared with four mainstream neural network models, including: a BP neural network optimized by a genetic algorithm (GA-BP), a generalized regression neural network (GRNN), a neural network based on the Northern Eagle optimization algorithm (NGO neural network), and a neural network based on the Whale optimization algorithm (WOA neural network).

[0152] See attached document Figure 3 This figure shows the distribution of prediction errors for each model on the test set. The horizontal axis represents different prediction models, and the vertical axis represents the percentage of error. The orange bars correspond to the mean absolute percentage error (MASE). The blue bars correspond to the maximum relative error ( ).

[0153] use The formula calculates the mean absolute percentage error, which is used to assess the overall deviation of the model's predicted values. The formula is:

[0154] ;

[0155] use The formula calculates the maximum relative error, which is used to evaluate the model's stability under extreme weather conditions. The formula is:

[0156] ;

[0157] In the formula, The total number of test samples; For the first Actual meteorological monitoring values ​​at each moment; For the first The model's predicted value at each time point.

[0158] According to the appendix Figure 3 Data distribution analysis shows that:

[0159] The WOA neural network model had the highest MAPE error, reaching approximately 9.7%, indicating that its overall prediction accuracy was low.

[0160] Although the MAPE error of the NGO neural network model was controlled at around 7.7%, its MAX error was as high as 9.9%, which was the highest among all the comparison models, indicating that the model has great instability when dealing with meteorological abrupt changes.

[0161] The Generalized Regressive Neural Network (GRNN) has a MAPE error of approximately 8.2% and a MAX error of approximately 8%, indicating moderate predictive performance.

[0162] The GA-BP model has a MAPE error of approximately 7% and a MAX error of approximately 6.5%, which is an improvement compared to the unfused single model.

[0163] In comparison, the fusion prediction model proposed in this invention performs best in both metrics, with a MAPE error of only 5.2% and a MAX error controlled at 5.8%. This indicates that by establishing a three-dimensional dynamic weighting system based on DSI, independence weighting, and entropy weighting, weights can be automatically allocated according to the model performance within different forecast lead times, thereby reducing the average prediction deviation and significantly suppressing the fluctuation of the maximum error.

[0164] See attached document Figure 4 The figure shows the root mean square error (RMSE) of different models. The comparison results are shown. The horizontal axis represents the model name, and the vertical axis represents... Numerical value.

[0165] use The formula calculates the root mean square error to measure the dispersion of the deviation between the predicted and the true values. The formula is:

[0166] ;

[0167] In the formula, This represents the number of test samples; These are actual monitored values; These are predicted values.

[0168] From the appendix Figure 4 It can be known that:

[0169] Generalized Regressive Neural Network (GRNN) The value is the highest, close to 1.9, indicating that its error dispersion is large in the fitting process of non-stationary meteorological series;

[0170] WOA neural networks and NGO neural networks They are approximately 1.4 and 1.25 respectively;

[0171] GA-BP model Approximately 1.05;

[0172] The model of this invention The value is the lowest, at approximately 0.95, making it the only model among all the comparison models with a value below 1.0.

[0173] The experimental results demonstrate that this invention optimizes the parameters of the BP and Generalized Regression Neural Network (GRNN) using the SWD and MUL algorithms respectively, and combines the generalization advantage of ELM to construct a differentiated group of sub-models. The fused model maintains high robustness while significantly reducing the root mean square error, providing a high-precision benchmark for subsequent residual signal extraction. This ensures that the LSTM network can accurately separate the feature changes caused by icing from the residuals, achieving accurate prediction of icing risk levels.

Claims

1. An icing risk early warning method based on multi-model fusion and residual time series feature analysis, characterized in that, The method comprises the following steps: S1, real-time collection of meteorological data of a target line area as actual observation values, identification and elimination of outliers in the actual observation values, standardization operation, and construction of an initial data set; wherein the meteorological data includes environmental temperature, daily minimum temperature, relative humidity, saturated humidity, absolute humidity, and wind speed; S2, evaluation and calculation of meteorological features in the initial data set by using a TEROL algorithm, screening of a preset number of meteorological features with high weight value ranking, and construction of a key meteorological feature data set; S3, input of the key meteorological feature data set into SWD-BP, MUL-GRNN, and ELM meteorological prediction models for parallel processing, output of prediction results of each sub-model, and calculation of a final meteorological prediction value based on a dynamic weight combination strategy; S4, calculation of a difference between the final meteorological prediction value and the actual observation value to generate a residual signal, extraction of time domain features of the residual signal, processing of the time domain features by using a sliding window algorithm, and construction of a residual feature matrix reflecting a dynamic change law of the residual; S5, input of the residual feature matrix into a pre-trained LSTM sequence classification model, capture of time sequence dependence in the residual feature matrix, and output of a current icing risk level result and warning information.

2. The icing risk early warning method based on multi-model fusion and residual time series feature analysis according to claim 1, characterized in that, In the S1 step, the specific process of identifying and eliminating outliers in the actual observation values and performing standardization operation comprises: Preliminary screening of the actual observation values based on the Laplace criterion, calculation of the average value and the standard deviation, and marking of data exceeding the set threshold as suspected outliers; Secondary discrimination of the suspected outliers, calculation of data change rates before and after the suspected outliers appear, and comparison with historical same-period meteorological extreme value distribution, if the suspected outliers meet the meteorological physical evolution law and are within the historical extreme value tolerance range, the suspected outliers are retained as true extreme values, otherwise, the suspected outliers are determined as false values; Elimination of the false values from the data sequence, and normalization processing of the remaining data after elimination to generate the initial data set.

3. The icing risk early warning method based on multi-model fusion and residual time series feature analysis according to claim 1, characterized in that, The S2 step specifically comprises: Random extraction of samples from the initial data set, and finding of correct neighbors in similar samples and wrong neighbors in different samples; Calculation of normalized distances of the samples and the correct neighbors and the samples and the wrong neighbors in each meteorological feature dimension; Iterative updating of weight values of each meteorological feature by using the normalized distances until a preset maximum iteration number is reached to obtain final weights of each feature; Sorting of the final weights, selection of a preset number of features with high weight value ranking as key features, and construction of the key meteorological feature data set by using the key features.

4. The icing risk early warning method based on multi-model fusion and residual time series feature analysis according to claim 1, characterized in that, In the S3 step, the process of constructing the SWD-BP and MUL-GRNN meteorological prediction models comprises: The SWD-BP weather prediction model is constructed by establishing a three-layer architecture BP neural network, encoding initial connection weights and threshold values of the BP neural network into position vectors, performing global iterative optimization on the position vectors by using a SWD algorithm, assigning the decoded optimal parameters to the BP neural network, and constructing the SWD-BP weather prediction model. The MUL-GRNN weather prediction model is constructed by establishing a generalized regression neural network GRNN, defining a variable matrix for representing a black ant colony state, performing cyclic optimization on a smoothing parameter SPREAD of a mode layer Gaussian function in the generalized regression neural network GRNN by using a MUL algorithm, substituting the determined optimal SPREAD parameter into the generalized regression neural network GRNN, and constructing the MUL-GRNN weather prediction model.

5. The icing risk early warning method based on multi-model fusion and residual time series feature analysis according to claim 4, characterized in that, In the SWD algorithm, a SWD parameter optimization formula is used to calculate parameter values of the next iteration, in the MUL algorithm, a roulette wheel update formula is used to preliminarily update variable values in the variable matrix, a tunnel exchange update formula is used to update material exchange between black ant colonies, and a parameter adjustment formula is used to calculate a tunnel existence probability and a fulfillment distance rate for balancing global and local search capabilities.

6. The ice-coating risk early warning method based on multi-model fusion and residual time-series feature analysis according to claim 1, characterized in that, In the S3 step, the specific steps of calculating the final weather prediction value based on a dynamic weight combination strategy include: obtaining prediction results of each sub-model after de-normalization, and dividing the prediction results into an initial group and a test group according to a time sequence; based on the test group data, counting distribution frequencies of prediction values of each sub-model above and below a set threshold value, calculating DSI weights based on prediction frequencies; performing error correlation analysis on the initial group data and weather live data of the previous day, calculating initial objective weights based on an independence weight method; for the test group data, evaluating information entropy values of each sub-model within a future forecast time limit, calculating objective weights based on an entropy weight method; combining the DSI weights based on prediction frequencies, the initial objective weights based on the independence weight method, and the objective weights based on the entropy weight method, constructing a combination weight matrix, using the combination weight matrix to perform weighted summation on prediction results of each sub-model, and outputting the final weather prediction value.

7. The icing risk early warning method based on multi-model fusion and residual time series feature analysis according to claim 6, characterized in that, In the process of calculating weights, the DSI weights based on prediction frequencies are calculated by using a DSI weight formula, the initial objective weights based on the independence weight method are calculated by using a negative correlation coefficient calculation formula and an initial objective weight formula, and the objective weights based on the entropy weight method are calculated by using an entropy weight method formula; the combination weight matrix is calculated by using a combination weight formula.

8. The ice-coating risk early warning method based on multi-model fusion and residual time-series feature analysis according to claim 1, characterized in that, In the S4 step, the time domain features of the residual error signal are calculated by using the following formulas: the arithmetic mean of the residual error signal is calculated by using a mean formula; the maximum value of the residual error signal is calculated by using a peak value formula; the difference between the maximum value and the minimum value of the residual error signal is calculated by using a peak-to-peak value formula; the total energy of the residual error signal is calculated by using a square sum formula; the square root amplitude of the residual error signal is calculated by using a root mean square amplitude formula.

9. The ice-coating risk early warning method based on multi-model fusion and residual time-series feature analysis according to claim 1, characterized in that, The training process of the LSTM sequence classification model in the S5 step includes: A loss function comprising a regularization term is constructed, which is composed of a cross-entropy error function and an L2 regularization term, and the transpose of a weight vector and half of the product of the weight vector are taken as the L2 regularization term; An improved random momentum gradient descent strategy is used to iteratively update the learnable parameters of the LSTM sequence classification model, and a random momentum parameter update formula is used to calculate the parameters of the next iteration, which introduces a random disturbance factor in the momentum term.

10. The ice-coating risk early warning method based on multi-model fusion and residual time-series feature analysis according to claim 1, characterized in that, In the S5 step, the output icing risk level is determined according to the equivalent icing thickness is divided into four levels: First risk, corresponding to no icing or a small amount of icing, ; Secondary risk, corresponding to light icing, ; Level 3 risk, corresponding to moderate icing, ; Quaternary risk, corresponding to heavy icing, .

Images