Line icing monitoring and early warning method and system based on neural network
Through the neural network-based line ice-covered monitoring and early warning method, combined with multi-source heterogeneous data and hidden Markov random field model, the problems of insufficient data fusion and insufficient flexibility of response strategies in the existing technology are solved, and high-precision The identification and early warning of line ice-covered risks ensure the safe and stable operation of distribution lines.
Patent Information
- Application Number
- CN202410668690.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2024-11-22
AI Technical Summary
The existing line ice-cover monitoring technology has problems such as insufficient data fusion, insufficient risk identification accuracy and lack of flexibility in response strategies, which makes it difficult to meet actual needs for early warning information timeliness and accuracy.
Using a neural network-based line ice-covered monitoring and early warning method, the multi-source heterogeneous data is collected in real time by deploying sensor arrays, and a hierarchical SVM classifier integration model and Bayesian network are constructed, combining the hidden Markov random field model to identify and predict Ice-covering risks, generate regional ice-covering warning signals, and formulate differentiated response strategies.
It improves the accuracy of risk identification and timeliness of early warning, can promptly detect the risk of line ice covering, accurately evaluate the degree of risk and impact range, and provides a basis for formulating differentiated response strategies to ensure the safe and stable operation of distribution lines.
Smart Images

Figure CN119005669A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power distribution line icing monitoring, and in particular to a neural network-based line icing monitoring and early warning method and system. Background Art
[0002] Icing of distribution lines is one of the main threats to the safe and stable operation of power systems. Icing can lead to accidents such as increased line load, flashover of insulators, and collapse of towers, causing large-scale power outages, resulting in serious economic losses and social impacts. Therefore, real-time monitoring and accurate early warning of line icing are crucial to ensure the safe and reliable operation of power systems. In recent years, with the rapid development of sensor technology, communication technology, and artificial intelligence technology, line icing monitoring and early warning technology has made significant progress. Traditional icing monitoring mainly relies on manual inspections, which is inefficient and poses safety hazards. At present, the more common monitoring method is to use optical, microwave and other sensors to collect line icing data and combine meteorological information for early warning. However, these methods often rely on a single data source, have limited perception capabilities for complex icing environments, and the accuracy and timeliness of early warnings are difficult to meet actual needs.
[0003] However, existing technologies still have many deficiencies in data fusion and risk identification. For example, the monitoring method based on single sensor data has limited accuracy and is difficult to fully reflect the actual situation of the line; when traditional statistical analysis methods process multi-source heterogeneous data, they cannot fully explore the complex associations between data, resulting in insufficient accuracy in risk identification. At the same time, existing risk assessment methods lack a dynamic response mechanism and cannot adjust early warning strategies based on real-time monitoring data, which can easily lead to delayed or too frequent early warning information, affecting the normal operation of the power grid. Summary of the invention
[0004] In view of the above problems, the present invention is proposed;
[0005] Therefore, the technical problem solved by the present invention is: how to provide a line icing monitoring and early warning method and system based on a neural network to isolate the existing icing monitoring technology from problems such as insufficient data fusion, insufficient risk identification accuracy, and lack of flexibility in response strategies.
[0006] To solve the above technical problems, the present invention provides the following technical solutions: a neural network-based line icing monitoring and early warning method, comprising: deploying a sensor array along the line to collect multi-source heterogeneous data around the line in real time; constructing a neural network-based line icing monitoring and early warning model based on the multi-source heterogeneous data, and identifying and predicting the line icing risk through the line icing monitoring and early warning model; when the line icing monitoring and early warning model predicts that there is an icing risk on the line, the degree and impact range of the icing risk are evaluated in combination with the geographical location information of the line, and a regionalized icing early warning signal is generated; a graded response strategy is formulated based on the degree of the icing risk; based on the icing early warning signal, a differentiated response strategy is formulated for different risk areas; according to the differentiated response strategy and the graded response strategy, the icing monitoring and early warning of the line is completed.
[0007] As a preferred solution of the neural network-based line icing monitoring and early warning method described in the present invention, the deployment strategy of the sensor array along the line includes multi-level and node clustering deployment of sensor nodes along the entire line, and setting the sensor node density according to the complexity of the line direction and terrain conditions; the multi-source heterogeneous data includes meteorological data, line monitoring data, line body monitoring data, historical icing event data, drone remote sensing data, artificial visual recognition data, and satellite remote sensing icing recognition data.
[0008] As a preferred solution of the neural network-based line icing monitoring and early warning method described in the present invention, wherein: the construction process of the line icing monitoring and early warning model includes the following steps: constructing a hierarchical SVM classifier integration model; wherein the hierarchical SVM classifier integration model includes a lower-layer SVM classifier and an upper-layer SVM classifier, the lower-layer SVM classifier is used to process a single data source and extract local patterns, and the upper-layer SVM classifier is used to fuse the output of the lower-layer SVM classifier to form an overall criterion; the feature mapping kernel function of the lower-layer SVM classifier is designed according to the multi-source heterogeneous data; the upper-layer SVM classifier uses a tensor kernel function to fuse the output of the lower-layer SVM classifier; constructing an initial Bayesian network structure based on domain knowledge, specifically defining observed variable nodes, hidden state variable nodes and target variable nodes, and determining the topological structure between the observed variable nodes, hidden state variable nodes and target variable nodes according to the causal relationship between the input data and expert knowledge;
[0009] Introduce prior probability distribution and learn network parameters, specifically, set prior probability distribution, use Bayesian parameter learning algorithm, combine prior probability distribution and training data, jointly estimate network parameters, and iterate parameter learning;
[0010] An SVM-Bayes hybrid training strategy is designed. Specifically, based on the output of the Bayesian network structure, the expected output distribution of each SVM classifier is calculated as a supervision signal, and the SVM parameters are optimized by gradient descent. The optimized SVM output is input back into the Bayesian network for iterative training.
[0011] As a preferred solution of the line icing monitoring and early warning method based on neural network described in the present invention, wherein: the lower layer SVM classifier includes a meteorological SVM classifier, a line monitoring SVM classifier, a line body monitoring SVM classifier, a historical icing event SVM classifier, an unmanned aerial vehicle remote sensing SVM classifier, an artificial visual recognition SVM classifier, and a satellite remote sensing icing recognition SVM classifier; the upper layer SVM classifier adopts a tensor kernel function to fuse the output of the lower layer SVM classifier, and comprises the following steps:
[0012] Assume that the output feature maps of the 7 lower-level SVM classifiers are:
[0013] φ(x)=[φ1(x),φ2(x),...,φ7(x)] T ;
[0014] Among them, φ(x) is the feature mapping output of the lower layer SVM classifier, φ1(x) is the feature mapping output of the first lower layer SVM classifier; x and y represent the sample data input to the upper layer SVM classifier.
[0015] The upper layer SVM uses the tensor product kernel function:
[0016]
[0017] in, is the product of two vectors φ(x) and φ(y); n is an integer hyperparameter.
[0018] When n=1, the tensor product kernel is equivalent to the linear kernel:
[0019]
[0020] When n>1, the tensor product kernel effectively captures the high-order statistical laws and correlations between multi-source features.
[0021] As a preferred solution of the neural network-based line icing monitoring and early warning method described in the present invention, during online monitoring, the multi-source heterogeneous data collected in real time are graded and classified, specifically, the lower-level SVM first classifies the source data and outputs it to the Bayesian network; the upper-level SVM makes an overall judgment on the line icing risk based on the Bayesian network probability reasoning output.
[0022] As a preferred embodiment of the neural network-based line icing monitoring and early warning method described in the present invention, the line icing risk includes low risk, medium risk and high risk; when the expected value or mean of the hidden state variable is less than the first preset threshold, and when the variance of the hidden state variable is less than the third preset threshold, the line icing is judged to be a low risk; when the first preset threshold ≤ the expected value or mean of the hidden state variable < the second preset threshold, or when the third preset threshold ≤ the variance of the hidden state variable < the fourth preset threshold, the line icing is judged to be a medium risk; when the variance of the hidden state variable is ≥ the second preset threshold, or when the variance of the hidden state variable is ≥ the fourth preset threshold, it is judged to be a low risk; wherein, the first preset threshold is less than the second preset threshold, and the third preset threshold is less than the fourth preset threshold.
[0023] As a preferred solution of the neural network-based line icing monitoring and early warning method described in the present invention, when the line icing monitoring and early warning model predicts that there is an icing risk on the line, the degree and impact range of the icing risk are evaluated in combination with the geographical location information of the line, and a regionalized icing early warning signal is generated, including the following steps: introducing a hidden Markov random field HMRF model, combining the geographical location information of the line with the icing risk level, and estimating the regional impact range and spatial distribution of icing; solving the state transition probability of the HMRF model by distance weighting of the Gaussian kernel function; combining the local evidence field of the HMRF model, unsupervised clustering of the regional icing distribution is performed by a non-parametric confidence density propagation algorithm to determine the spatial range of the icing impact and the risk level distribution within the range; vectorizing the clustering results to generate a regionalized icing early warning signal.
[0024] To further solve the above technical problems, the present invention provides the following technical solutions: a distribution line icing monitoring and early warning system based on a neural network, comprising: a data acquisition module, used to deploy a sensor array along the line to collect multi-source heterogeneous data in real time; a model construction module, used to construct a line icing monitoring and early warning model based on a neural network according to the multi-source heterogeneous data; a risk identification module, used to identify and predict the line icing risk through the line icing monitoring and early warning model, and when the icing risk is predicted, the icing risk degree and impact range are evaluated; an early warning generation module, used to generate a regionalized icing early warning signal based on the evaluation result of the risk identification module.
[0025] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the line icing monitoring and early warning method based on a neural network as described above are implemented.
[0026] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the line icing monitoring and early warning method based on a neural network as described above are implemented.
[0027] Beneficial effects of the present invention: The present invention proposes a line icing monitoring and early warning method and system based on a neural network. The scheme deploys a sensor array along the line to collect multi-source heterogeneous data in real time, and constructs a line icing monitoring and early warning model based on a neural network based on these data, and uses the model to identify and predict the line icing risk. When it is predicted that there is an icing risk, the icing risk level and impact range are evaluated in combination with the line's geographical location information, and a regionalized icing early warning signal is generated. The scheme integrates multi-source heterogeneous data to improve the accuracy of risk identification; adopts a hybrid training strategy that combines a hierarchical SVM classifier integration model with a Bayesian network to enhance the generalization ability of the model; introduces a hidden Markov random field model to evaluate the icing impact range, generates a regionalized early warning signal, and improves the accuracy and timeliness of the early warning. The overall technical solution can timely discover the line icing risk, accurately evaluate the risk level and impact range, and provide a basis for formulating differentiated response strategies, thereby ensuring the safe and stable operation of the distribution lines. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 work.
[0029] Figure 1 It is an overall flow chart of a line icing monitoring and early warning method based on a neural network of the present invention;
[0030] Figure 2 A diagram of a computer device according to the present invention. DETAILED DESCRIPTION
[0031] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings in the specification. It is obvious that the described embodiments are part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments in the present invention, all other embodiments obtained by ordinary persons in the art without making creative work should fall within the scope of protection of the present invention.
[0032] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0033] Example 1
[0034] Reference Figure 1 , which is an embodiment of the present invention, provides a line icing monitoring and early warning method based on a neural network, comprising:
[0035] S1: Deploy sensor arrays along the route to collect multi-source heterogeneous data around the route in real time.
[0036] S1.1: Determine the sensor deployment strategy along the line. The layout of the sensor array needs to meet the following aspects: a. Comprehensive coverage: Sensor nodes are evenly deployed along the entire transmission line to ensure that the entire line is monitored without omission. b. Moderate node density: According to the complexity of the line direction, terrain conditions, etc., the sensor node density is reasonably set to avoid being too dense or too sparse. c. Multi-level deployment: In addition to deployment along the ground, sensors can be added to key locations such as line towers and insulators to form multi-level three-dimensional perception. d. Node clustering: In key areas such as ice-prone areas and line turning points, sensor node clusters are deployed to improve local monitoring accuracy.
[0037] S1.2: Collect multi-source heterogeneous data around the line in real time through the deployed sensor array.
[0038] Specifically, multi-source heterogeneous data include: a. Meteorological data: temperature, humidity, wind speed, wind direction, precipitation, visibility, etc., which can reflect changes in weather conditions and directly affect the formation and melting of line icing. b. Line monitoring data: Using the monitoring video equipment installed along the line to collect line real-time image data, you can visually observe whether the line is covered with ice. d. Line body monitoring data: online temperature, load, vibration, etc. of key parts such as transmission line gaps and insulators can reflect the icing status. e. Historical icing event data: Cases of icing events that have occurred on the same line can provide references for model training. f. Drone remote sensing data: Using drones to regularly collect high-precision image data can more accurately identify the icing status. g. Artificial visual recognition data: Analyze monitoring images through AI visual algorithms and output line icing status analysis results. h. Satellite remote sensing icing recognition data: Use satellite remote sensing technology to identify icing in a large area.
[0039] S2: Construct a neural network-based line icing monitoring and early warning model based on multi-source heterogeneous data, and identify and predict line icing risks through the line icing monitoring and early warning model.
[0040] Preferably, the present invention constructs a line icing monitoring and early warning model by combining a support vector machine (SVM) and a Bayesian network.
[0041] Support vector machines (SVMs) are widely used in pattern recognition and data classification problems and have good generalization capabilities. Therefore, the present invention studies and designs feature mapping kernel functions to map multi-source heterogeneous data into a high-dimensional feature space, so that the ice-covered and non-ice-covered states of the line can be more separable in the space. At the same time, in view of the spatiotemporal correlation of the data, a dynamic kernel function is designed to capture the temporal and spatial constraint information.
[0042] In addition, since icing of transmission lines involves multiple influencing factors, such as meteorological conditions, geographical environment, line status, etc., these factors have complex correlations and conditional independence, and the Bayesian network can effectively model the dependencies between these random variables, express the conditional independence assumptions between variables through a graphical model, and perform probabilistic reasoning on the risk of line icing. The Bayesian network uses a graphical model to represent the conditional independence between random variables, and can handle uncertain data well. The present invention constructs the initial Bayesian network structure and prior probability based on domain knowledge, and then automatically adjusts the network parameters through machine learning to form a joint probability model that integrates multi-source data to reason about the risk of line icing.
[0043] Therefore, when constructing a line icing monitoring and early warning model, the present invention complements the advantages of SVM and Bayesian network: SVM is used to discriminate local patterns (such as meteorological conditions, video surveillance data, etc.) and use them as input nodes of the Bayesian network; SVM output and other data sources (such as geographic location, line status, etc.) are integrated in the Bayesian network to capture the conditional independence relationship between data; and the Bayesian network is used to perform probabilistic reasoning and decision-making on the overall icing risk. This combined model architecture maintains modular characteristics, and each module can be easily adjusted and replaced, such as replacing the SVM kernel function, modifying the Bayesian network structure, etc., and has good flexibility and scalability.
[0044] S2.1: Construct a hierarchical SVM classifier ensemble model based on multi-source heterogeneous data.
[0045] Specifically, for the multi-source heterogeneous data in step S1, multiple SVM classifiers are trained to form a hierarchical and ordered structure. The lower-level SVM classifier focuses on processing a single data source and extracting local patterns; the upper-level SVM classifier fuses the outputs of the lower-level SVM classifiers to form an overall criterion.
[0046] It should be noted that the lower-level SVM classifiers include 7, which correspond to the meteorological data, line monitoring data, line body monitoring data, historical icing event data, UAV remote sensing data, artificial visual recognition data, and satellite remote sensing icing recognition data in step S1, respectively. They are meteorological SVM classifier, line monitoring SVM classifier, line body monitoring SVM classifier, historical icing event SVM classifier, UAV remote sensing SVM classifier, artificial visual recognition SVM classifier and satellite remote sensing icing recognition SVM classifier.
[0047] S2.2: Design the feature mapping kernel function of the lower-level SVM classifier based on multi-source heterogeneous data.
[0048] Specifically, the feature mapping kernel function of the meteorological SVM classifier adopts the Gaussian kernel function; the line monitoring SVM classifier adopts the additive combination of the linear kernel and the Gaussian kernel; the line body monitoring SVM classifier adopts the polynomial kernel function; the historical icing event SVM classifier adopts the hyperbolic tangent kernel function; the UAV remote sensing SVM classifier adopts the Gaussian sequence kernel function; the artificial vision recognition SVM classifier adopts the generalized kernel function; and the satellite remote sensing icing recognition SVM classifier adopts the radial basis function kernel.
[0049] S2.3: The upper layer SVM classifier uses the tensor kernel function to fuse the output of the lower layer SVM classifier.
[0050] Specifically, the output feature maps of the 7 lower-level SVM classifiers are assumed to be:
[0051] φ(x)=[φ1(x),φ2(x),...,φ7(x)] T ;
[0052] Among them, φ(x) is the feature map output of the lower layer SVM classifier, φ i (x) is the feature map output of the i-th lower-level SVM classifier; x and y represent the sample data input to the upper-level SVM classifier.
[0053] Then, the upper SVM uses the tensor product kernel function as follows:
[0054]
[0055] in, It is the product of two vectors φ(x) and φ(y); n is an integer hyperparameter;
[0056] When n=1, the tensor product kernel is equivalent to the linear kernel:
[0057]
[0058] When n>1, the tensor product kernel can effectively capture the high-order statistical laws and correlations between multi-source features.
[0059] Usually, the value of n ranges from [1,5]. A larger value of n will increase the complexity and computational cost of the kernel function. The optimal value of n can be selected from the candidate set {1,2,3,4,5} by grid search and cross validation.
[0060] In the implementation process, the upper SVM outputs φ of the lower 7 SVMs i (x) Perform tensor product operations according to the vector form φ(x) to fuse into high-order features, and then obtain the final ice cover recognition model based on the tensor product kernel K(x,y) training to perform overall criterion fusion on heterogeneous multi-source data.
[0061] S2.4: Construct an initial Bayesian network structure based on domain knowledge.
[0062] Specifically, the following three types of node variables are defined:
[0063] Observed variable node: includes the outputs of the 7 lower-level SVM classifiers in step S2.1, as well as other observable line monitoring data.
[0064] Hidden state variable nodes: including hidden variables such as regional icing degree and meteorological conditions that are difficult to observe directly but affect the icing risk.
[0065] Target variable node: Overall icing risk assessment.
[0066] Based on the causal relationship between the input data and expert knowledge, the topological structure between the three types of variable nodes is determined:
[0067] a) Lower layer SVM output node φ i (x) as an observed variable node, pointing to the corresponding hidden state variable node, such as the degree of regional ice coverage;
[0068] b) Other observable line monitoring data nodes also point to corresponding hidden state variable nodes;
[0069] c) There can be dependencies between hidden state variable nodes, such as the degree of regional ice coverage being affected by hidden variables such as meteorological conditions;
[0070] d) All hidden state variable nodes point to the final target variable overall icing risk assessment node.
[0071] S2.5: Introduce prior probability distribution and learn network parameters.
[0072] Specifically, set the prior probability distribution. For the observed variable node, that is, the output node φ of the lower layer SVM classifier i(x) and other observable line monitoring data nodes, the prior probability distribution can be set according to historical data; while for hidden state variable nodes, such as regional ice coverage, meteorological conditions, etc., reasonable prior probability distribution is set based on domain expert knowledge.
[0073] Furthermore, parameter learning is performed. Using the Bayesian parameter learning algorithm, combined with the above prior probability distribution and training data, the network parameters are jointly estimated:
[0074]
[0075] Where Pa(X) is the parent node set of node X; P0(X|Pa(X)) is the prior probability distribution of X given Pa(X); P(D|X,Pa(X)) is the data likelihood of X given Pa(X) under the observed data D; is the conditional probability distribution parameter that needs to be estimated.
[0076] It should be noted that different distribution functions and parameter learning methods can be used according to the different properties of variable nodes: for discrete variable nodes, it is generally assumed to obey multinomial distribution, and the expectation-maximization algorithm can be used to estimate parameters; for continuous variable nodes, it can be assumed to obey Gaussian distribution, t-distribution, etc., and the parameters can be estimated using techniques such as Bayesian linear regression.
[0077] Furthermore, the parameter learning is iterated. In the Bayesian network parameter learning process, the following steps can be iteratively performed:
[0078] Use the current network parameters to perform probabilistic reasoning on hidden variable nodes (such as regional ice coverage) to obtain their distribution;
[0079] Use the inference results as prior knowledge to re-estimate the network parameters;
[0080] Repeat the above two steps until convergence.
[0081] This iterative approach of parameter learning and probabilistic reasoning can fully integrate prior knowledge and data information and continuously improve network parameters.
[0082] S2.6: Design a SVM-Bayes hybrid training strategy.
[0083] Specifically, the expected output of SVM is calculated based on the Bayesian network, and the Bayesian network parameters learned by S2.5 are used to perform probabilistic reasoning on the hidden state variable nodes to obtain their posterior probability distribution:
[0084] P(H|O,M)∝P(O|H,M)P(H|M);
[0085] Among them, H is the hidden state variable node; H is the observed variable node, including other monitoring data; M is the known network structure and parameters.
[0086] According to P(H|O,M) and hidden variables H and the lower layer SVM output φ i The conditional probability distribution of (x) is used to calculate the expected output distribution of each SVM classifier:
[0087] E[φ i (x)|O,M]=∫φ i (x)P(φ i (x)|H,O,M)dH;
[0088] Furthermore, based on the expected output, the SVM parameters are optimized and the above expected output distribution E[φ i (x)|O,M] is used as a supervisory signal to perform gradient descent optimization on the parameters of each lower-level SVM classifier:
[0089]
[0090] in, is the loss function, such as mean square loss or cross entropy loss; θ i To represent the parameter set of i lower-level SVM classifiers, its optimal value needs to be determined through optimization φ i (x,θ i ) is the i-th lower layer SVM classifier under the input sample x and the current parameter θ i The feature map output under i (x)|O,M] represents the expected output distribution of the i-th SVM classifier calculated based on the observed data O and the known Bayesian network structure and parameters M; To express the need to find a loss function The optimal parameters for minimization
[0091] By solving the above optimization problem, we find the optimal parameters that minimize the difference between the actual output of the i-th SVM classifier and the expected output distribution calculated based on the Bayesian network.
[0092] Furthermore, SVM-Bayes is trained iteratively.
[0093] Initialization: Use domain knowledge and limited labeled data to perform feed-forward training on SVM and Bayes network to obtain initial parameters.
[0094] Iterative training:
[0095] 1) Using the current Bayesian network parameters, calculate the expected output of SVM;
[0096] 2) Based on the expected output, optimize the parameters of each SVM;
[0097] 3) Input the optimized SVM output back into the Bayes network;
[0098] 4) Relearn the Bayes network parameters;
[0099] 5) Repeat the above steps until convergence.
[0100] S2.7: During online monitoring, the multi-source heterogeneous data collected in real time are classified and judged. The lower-level SVM first classifies the source data and outputs it to the Bayesian network; the upper-level SVM makes an overall judgment on the line icing risk based on the Bayesian network probability reasoning output.
[0101] Specifically, the Bayesian network infers the posterior probability distribution of the hidden state variable (such as the degree of regional icing) P(H|O,M). According to P(H|O,M), the expected value, variance and other statistics of the hidden state variable $H$ can be obtained as indicators of the degree of icing risk. These statistics are compared with the pre-set icing risk classification threshold to determine the current icing risk level.
[0102] Furthermore, line icing risks include low risk, medium risk and high risk.
[0103] Let the expected value or mean of the hidden state variable be E[H], which can be regarded as a point estimate of the hidden state variable H (ice coverage). When evaluating ice risk, a higher E[H] value indicates that the ice coverage in the region is more serious based on the current observation data and model. Therefore, by comparing E[H] with the classification threshold, the current ice risk level can be determined. The expression of E[H] is
[0104] E[H]=∫HP(H|O,M)dH.
[0105] At the same time, let the variance of the hidden state variable H (such as the degree of regional ice coverage) be Var[H]. The variance reflects the degree of discreteness of the distribution of a random variable value and is used to measure the volatility or uncertainty of the random variable value. The expression of Var[H] is:
[0106] Var[H]=E[(HE[H]) 2 ];
[0107] Specifically, Var[H] is the second-order moment of H computed from the posterior probability distribution of H, P(H|O,M):
[0108] Var[H]=∫(HE[H]) 2 p(H|O,M)dH;
[0109] A larger Var[H] value indicates that based on the current observed data and model, there is greater uncertainty in H and a wider range of possible values. Conversely, a smaller Var[H] value indicates that the posterior distribution of H is more concentrated and the uncertainty is smaller.
[0110] When assessing the risk of icing, in addition to considering the value of E[H], it is also necessary to consider Var[H]:
[0111] When E[H] is large and Var[H] is small, it indicates that the icing risk is high and the certainty is strong.
[0112] When E[H] is large and Var[H] is also large, it indicates that although the ice cover is expected to be heavy, there is a large uncertainty.
[0113] Therefore, Var[H] is compared with the corresponding threshold to more comprehensively evaluate the severity and certainty of the icing risk and make corresponding risk level judgments and decisions.
[0114] Low risk: E[H] < τ l And Var[H]<σ l .
[0115] Medium risk: τ l ≤E[H]<τ h or l ≤Var[H]<σ h .
[0116] High risk: E[H] ≥ τ h or Var[H]≥σ h .
[0117] Wherein, the first preset threshold τ is set l , the second preset threshold τ h , the third preset threshold σ l , the fourth preset threshold σ h The relationship is: l <τ h , σ l <σ h .
[0118] The first preset threshold τ l is the lower limit of the threshold of the expected value E[H] of H, and the second preset threshold τ h It is the upper threshold of the expected value E[H] of H. The third preset threshold σ l is the lower limit of the variance Var[H] of H, and the fourth preset threshold σ h It is the upper threshold value of the variance Var[H] of H.
[0119] Preferably, this step realizes the dual constraints on the expected value and variance of the hidden state variable by setting four preset thresholds, thereby constructing a two-dimensional risk criterion. This criterion divides the risk space into three levels: low, medium, and high, making the risk assessment more meticulous and the transition smoother and more continuous. By combining the two statistics E[H] and Var[H], not only can the point estimate of the ice coverage degree be evaluated, but also the size of its related uncertainty can be quantified, avoiding the risk omission or misjudgment that may be caused by a single expected value criterion. Reasonable setting of the four thresholds can achieve a good balance between the misjudgment rate and the missed judgment rate, and improve the accuracy and reliability of risk judgment. This criterion embodies the integration of Bayesian estimation theory and domain expert knowledge. The former provides numerical support for the posterior distribution of H, and the latter provides empirical guidance for threshold setting. The two promote each other and ensure the scientificity and interpretability of the criterion. Finally, through this threshold criterion, the active, refined, and intelligent control of the line ice risk is realized, which greatly improves the safe and reliable operation level of the transmission line.
[0120] S3: When the line icing monitoring and early warning model predicts that there is an icing risk on the line, the geographical location information of the line is integrated to evaluate the impact range and risk level of the icing risk and generate a regionalized icing early warning signal.
[0121] Specifically, the Hidden Markov Random Field (HMRF) model is introduced to combine the geographical location information of the line with the ice risk level to estimate the regional impact range and spatial distribution of ice. The state transition probability of the HMRF model is solved by the distance weighting of the Gaussian kernel function. Combined with the local evidence field of the HMRF model, the regional ice distribution is unsupervisedly clustered by the non-parametric confidence density propagation algorithm to determine the spatial range of ice impact and the risk level distribution within the range. The clustering results are vectorized to generate a regionalized ice warning signal, and the regionalized ice warning signal is used to describe the geographical coordinates and shape of the impact range, as well as the distribution ratio of each risk level within the range.
[0122] In summary, the present invention proposes a line icing monitoring and early warning method and system based on neural network. The scheme deploys a sensor array along the line to collect multi-source heterogeneous data in real time, and constructs a line icing monitoring and early warning model based on neural network according to these data, and uses the model to identify and predict the line icing risk. When the risk of icing is predicted, the degree of icing risk and the scope of influence are evaluated in combination with the geographical location information of the line, and a regionalized icing early warning signal is generated. The scheme integrates multi-source heterogeneous data to improve the accuracy of risk identification; adopts a hybrid training strategy combining a hierarchical SVM classifier integration model and a Bayesian network to enhance the generalization ability of the model; introduces a hidden Markov random field model to evaluate the scope of icing influence, generates a regionalized early warning signal, and improves the accuracy and timeliness of the early warning. The overall technical scheme can timely discover the risk of line icing, accurately evaluate the degree of risk and the scope of influence, and provide a basis for formulating differentiated response strategies, thereby ensuring the safe and stable operation of distribution lines.
[0123] Example 2
[0124] Reference Figure 2 , is an embodiment of the present invention, and provides a distribution line icing monitoring and early warning system based on a neural network, including: a data acquisition module, used to deploy a sensor array along the line to collect multi-source heterogeneous data in real time; a model construction module, used to build a line icing monitoring and early warning model based on a neural network according to the multi-source heterogeneous data; a risk identification module, used to identify and predict the line icing risk through the line icing monitoring and early warning model, and when the icing risk is predicted, the icing risk degree and impact range are evaluated; an early warning generation module, used to generate a regionalized icing early warning signal based on the evaluation result of the risk identification module.
[0125] Example 3
[0126] Reference Figure 2 , is an embodiment of the present invention, which is different from the previous embodiment in that:
[0127] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0128] The logic and / or steps represented in the flowchart or otherwise described herein, for example, may be considered as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or in combination with such instruction execution systems, apparatuses or devices; for the purposes of this specification, "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, apparatus or device, or in combination with such instruction execution systems, apparatuses or devices;
[0129] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wirings (electronic devices), a portable computer disk case (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM); in addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory;
[0130] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof; in the above-mentioned embodiment, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system; for example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or a combination thereof: a discrete logic circuit having a logic gate circuit for implementing a logical function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.
[0131] Example 4
[0132] As an embodiment of the present invention, a line icing monitoring and early warning method based on a neural network is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.
[0133] The experiment considered four evaluation indicators: accuracy of icing risk identification, accuracy of icing impact range estimation, completeness of regional warning signal generation, and timeliness of warning response. It was compared with the existing statistical learning method based on single-source data (Benchmark 1) and the classifier fusion method based on multi-source data (Benchmark 2). The experimental data comes from the historical data set of icing events on a certain transmission line.
[0134] The evaluation indicators and comparison results are shown in the following table:
[0135] Table 1 Evaluation results
[0136] Evaluation indicators The present invention Benchmark1 Benchmark2 Accuracy of ice risk identification 0.928 0.793 0.867 Average error of impact range estimation (km) 3.21 8.94 6.37 Regionalized early warning signal completeness score 4.7 2.1 3.5 Average warning response delay (min) 17.6 32.4 26.3
[0137] As can be seen from Table 1, firstly, in terms of ice risk identification accuracy, the present invention improves risk identification capability by fusing multi-source heterogeneous data and using a neural network model, with an accuracy of 92.8%, significantly higher than the single-source statistical learning method (79.3%) and the traditional classifier fusion method (86.7%). Secondly, in terms of the accuracy of impact range estimation, the unconstrained clustering estimation method based on the hidden Markov random field model has an average error of only 3.21km, which is much lower than Benchmark 1 (8.94km) and Benchmark 2 (6.37km) based on the grid model. Then, the completeness of the regionalized warning signal is evaluated by scoring by domain experts. The vectorized warning signal generated by the present invention that describes the impact range and risk distribution has richer and more complete information (4.7 points), which is better than Benchmark 1 (2.1 points) and Benchmark 2 (3.5 points). Finally, in terms of the timeliness of early warning response, the end-to-end delay of the present invention is an average of 17.6 minutes, which is significantly improved compared with Benchmark 1 (32.4 minutes) and Benchmark 2 (26.3 minutes).
[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A line icing monitoring and early warning method based on neural network, characterized in that: include: Deploy sensor arrays along the route to collect multi-source heterogeneous data around the route in real time; Constructing a line icing monitoring and early warning model based on a neural network according to the multi-source heterogeneous data, and identifying and predicting line icing risks through the line icing monitoring and early warning model; When the line icing monitoring and early warning model predicts that there is an icing risk on the line, the degree and impact range of the icing risk are evaluated in combination with the geographical location information of the line, and a regionalized icing early warning signal is generated.
2. The line icing monitoring and early warning method based on a neural network according to claim 1, characterized in that: The deployment strategy of the sensor array along the route includes arranging sensor nodes in a multi-layered and node-clustered manner along the entire route, and setting the density of sensor nodes according to the complexity of the route and the terrain conditions; The multi-source heterogeneous data include meteorological data, line monitoring data, line body monitoring data, historical icing event data, drone remote sensing data, artificial visual recognition data, and satellite remote sensing icing recognition data.
3. The line icing monitoring and early warning method based on a neural network as claimed in claim 2 is characterized in that: The construction process of the line icing monitoring and early warning model includes the following steps: Constructing a hierarchical SVM classifier integration model; wherein the hierarchical SVM classifier integration model includes a lower-layer SVM classifier and an upper-layer SVM classifier, the lower-layer SVM classifier is used to process a single data source and extract local patterns, and the upper-layer SVM classifier is used to fuse the output of the lower-layer SVM classifier to form an overall criterion; Designing a feature mapping kernel function of the lower layer SVM classifier according to the multi-source heterogeneous data; The upper layer SVM classifier uses a tensor kernel function to fuse the output of the lower layer SVM classifier; Constructing a Bayesian network structure based on domain knowledge, specifically defining observation variable nodes, hidden state variable nodes and target variable nodes, and determining the topological structure among the observation variable nodes, hidden state variable nodes and target variable nodes according to the causal relationship between input data and expert knowledge; Introduce prior probability distribution and learn network parameters, specifically, set prior probability distribution, use Bayesian parameter learning algorithm, combine prior probability distribution and training data, jointly estimate network parameters, and iterate parameter learning; An SVM-Bayes hybrid training strategy is designed. Specifically, based on the output of the Bayesian network structure, the expected output distribution of each SVM classifier is calculated as a supervision signal, and the SVM parameters are optimized by gradient descent. The optimized SVM output is input back into the Bayesian network for iterative training.
4. The line icing monitoring and early warning method based on a neural network as claimed in claim 3 is characterized in that: The lower layer SVM classifiers include a meteorological SVM classifier, a line monitoring SVM classifier, a line body monitoring SVM classifier, a historical icing event SVM classifier, an unmanned aerial vehicle remote sensing SVM classifier, an artificial visual recognition SVM classifier, and a satellite remote sensing icing recognition SVM classifier; The upper layer SVM classifier uses a tensor kernel function to fuse the output of the lower layer SVM classifier, including the following steps: Assume that the output feature maps of the 7 lower-level SVM classifiers are: φ(x)=[φ1(x),φ2(x),...,φ7(x)] T ; Among them, φ(x) is the feature mapping output of the lower layer SVM classifier, φ1(x) is the feature mapping output of the first lower layer SVM classifier; x and y represent the sample data input to the upper layer SVM classifier; The upper layer SVM uses the tensor product kernel function: in, It is the product of two vectors φ(x) and φ(y); n is an integer hyperparameter; When n=1, the tensor product kernel is equivalent to the linear kernel: When n>1, the tensor product kernel effectively captures the high-order statistical laws and correlations between multi-source features.
5. The line icing monitoring and early warning method based on a neural network as claimed in claim 4 is characterized in that: During online monitoring, the multi-source heterogeneous data collected in real time are graded and classified. Specifically, the lower-level SVM first classifies the data from each source and outputs it to the Bayesian network. The upper-level SVM makes an overall judgment on the line icing risk based on the probabilistic reasoning output of the Bayesian network.
6. The line icing monitoring and early warning method based on a neural network as claimed in claim 5, characterized in that: The line icing risks include low risk, medium risk and high risk; When the expected value or mean value of the hidden state variable is less than the first preset threshold value, and when the variance of the hidden state variable is less than the third preset threshold value, it is determined that the line icing is of low risk; When the first preset threshold ≤ the expected value or mean of the hidden state variable < the second preset threshold, or when the third preset threshold ≤ the variance of the hidden state variable < the fourth preset threshold, it is determined that the line icing is of medium risk; When the variance of the hidden state variable is ≥ the second preset threshold, or when the variance of the hidden state variable is ≥ the fourth preset threshold, it is determined to be low risk; Among them, the first preset threshold is less than the second preset threshold, and the third preset threshold is less than the fourth preset threshold.
7. The line icing monitoring and early warning method based on a neural network according to claim 6, characterized in that: When the line icing monitoring and early warning model predicts that there is an icing risk on the line, the degree and impact range of the icing risk are evaluated in combination with the geographical location information of the line, and a regionalized icing early warning signal is generated, including the following steps: The Hidden Markov Random Field (HMRF) model is introduced to combine the geographical location information of the line with the icing risk level to estimate the regional impact range and spatial distribution of icing. The state transition probability of the HMRF model is solved by the distance weighting of the Gaussian kernel function; Combined with the local evidence field of the HMRF model, the regional ice distribution is clustered unsupervisedly using the non-parametric confidence density propagation algorithm to determine the spatial extent of the ice impact and the risk level distribution within the extent. The clustering results are vectorized to generate regionalized ice cover warning signals.
8. A system using the line icing monitoring and early warning method based on a neural network as claimed in any one of claims 1 to 7, characterized in that: include: Data acquisition module, used to deploy sensor arrays along the route to collect multi-source heterogeneous data in real time; A model building module is used to build a line icing monitoring and early warning model based on a neural network according to multi-source heterogeneous data; The risk identification module is used to identify and predict the line icing risk through the line icing monitoring and early warning model. When the icing risk is predicted, the icing risk degree and impact range are evaluated; The warning generation module is used to generate regional ice cover warning signals based on the evaluation results of the risk identification module.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the line icing monitoring and early warning method based on a neural network as described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the line icing monitoring and early warning method based on a neural network as described in any one of claims 1 to 7 are implemented.
Citation Information
Cited By
Model-data combined drive power transmission line icing state sensing and operation risk assessment method and device based on satellite internet
CN119646671A
Intelligent identification and evaluation method and system for external damage risk of power transmission channel
CN119671287A
Bank network risk control optimization method and system based on flow behavior analysis
CN120297500A
Bank network risk control optimization method and system based on traffic behavior analysis
CN120297500B
Agglomerate fog monitoring and early warning method and system based on multi-source optical information fusion
CN121257761A
