CN-121980914-A - Icing torsion inversion method, system, equipment and medium based on physical constraint neural network

CN121980914ACN 121980914 ACN121980914 ACN 121980914ACN-121980914-A

Abstract

The invention relates to the technical field of power system monitoring and discloses an icing torsion inversion method, system, equipment and medium based on a physical constraint neural network, which comprises the steps of collecting torsion angle data, environment parameter data and historical icing thickness data of a power transmission wire; the method comprises the steps of establishing a physical constraint model between a torsion angle and an icing thickness based on a wire mechanics principle, establishing a first neural network based on the physical constraint model, defining a composite loss function comprising a data driving loss function and the physical constraint loss function, training the first neural network by utilizing historical icing thickness data, optimizing network parameters through a back propagation algorithm to minimize the composite loss function, and inputting torsion angle data and environment parameter data into the trained first neural network for prediction to obtain a current icing thickness predicted value of a wire. According to the invention, the mechanical physical model is embedded into the neural network training process, so that accurate inversion from the wire torsion angle to the icing thickness is realized.

Inventors

WU YU
ZHENG SHUYI
LI ZHENGXIN
CAI DENGSHENG
WANG YOUJUN
YIN FANGHUI
ZHOU WENXUAN
LIAO YONGLI
HUANG JIE
Nie Xianglun
LI YI
ZHENG XIAOHU
LIU QING

Assignees

贵州电网有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251218

Claims (10)

1. An icing torsion inversion method based on a physical constraint neural network is characterized by comprising the following steps: Collecting torsion angle data, environmental parameter data and historical icing thickness data of a power transmission wire; establishing a physical constraint model between a torsion angle and the ice coating thickness based on a wire mechanics principle; constructing a first neural network based on the physical constraint model, and defining a composite loss function comprising a data-driven loss function and a physical constraint loss function; Training the first neural network by using the historical icing thickness data, and optimizing network parameters by a back propagation algorithm to minimize the composite loss function; And inputting the torsion angle data and the environmental parameter data into a trained first neural network for prediction to obtain the icing thickness predicted value of the current wire.
2. The method of ice coating torsion inversion based on a physical constraint neural network of claim 1, wherein constructing a first neural network based on the physical constraint model comprises: Determining the number of input layer nodes, the corresponding torsion angle and the environmental parameters of the first neural network; configuring a hidden layer structure of the first neural network by adopting a multi-layer full-connection network; and setting an output layer node, wherein the output layer is used for outputting the icing thickness predicted value.
3. A method of ice-on-torsion inversion based on a physically constrained neural network as claimed in claim 2, wherein defining a composite loss function comprising a data-driven loss function and a physically constrained loss function includes: the data driving loss function is used for calculating the mean square error loss between the icing thickness predicted value and the real icing thickness measured value; the physical constraint loss function is composed of a plurality of sub-items, wherein the first sub-item is used for constraining the icing thickness predicted value to meet a balance relation after being substituted into a torque balance equation, the second sub-item is used for constraining the predicted icing quality to accord with a physical rule, and the third sub-item is used for constraining the icing thickness predicted value to be within a reasonable physical boundary range; and distributing and adding overall weight coefficients to the data driving loss function and the physical constraint loss function to form the composite loss function.
4. A method of ice coating twist inversion based on a physically constrained neural network as claimed in claim 3, wherein said training said first neural network using said historical ice coating thickness data comprises: normalizing and preprocessing the collected torsion angle data, environment parameter data and historical icing thickness data, and dividing the preprocessed data into a training set and a verification set; Dividing the training set into a plurality of batches, inputting the batches into the first neural network, and optimizing parameters of the first neural network by adopting a back propagation algorithm with the aim of minimizing a composite loss function; And in the training process, evaluating the performance of the first neural network by using the verification set, stopping training and saving the network parameters after training when the loss of the verification set is not reduced for a plurality of training rounds.
5. The method for inverting icing torsion based on a physical constraint neural network according to claim 1, wherein the step of collecting torsion angle data, environmental parameter data and historical icing thickness data of the power transmission wire further comprises: when the acquired torsion angle data, environmental parameter data and historical icing thickness data are insufficient, random noise which accords with actual measurement error distribution is added into a physical model calculation result, and simulation data are generated to be used as supplement; The ratio of the simulation data to the measured data is controlled between 1:1 and 3:1.
6. The method for inverting ice coating torsion based on a physical constraint neural network according to claim 5, wherein the establishing a physical constraint model between torsion angle and ice coating thickness based on the principle of wire mechanics comprises: establishing a torque balance equation based on the balance relation between the torque generated by the ice coating load of the wire and the elastic recovery torque of the wire; Constructing an icing quality calculation model according to the icing density, the wire diameter and the icing thickness; an eccentric load model is constructed by considering eccentric moment caused by uneven distribution of ice coating in the circumferential direction of a wire; And combining the torque balance equation, the icing mass calculation model and the eccentric load model, and establishing a physical constraint model between the torsion angle and the icing thickness.
7. The method of ice-on-torsion inversion based on a physically constrained neural network of claim 6, further comprising: Continuously collecting new torsion angle data and corresponding icing thickness measurement data in the actual running process; Performing incremental training on the trained first neural network by periodically using the newly acquired data to update network parameters; Setting a model update triggering mechanism, and automatically triggering one-time parameter update of the first neural network when the accumulated new data quantity reaches a preset threshold value.
8. An icing torsion inversion system based on a physical constraint neural network, and an icing torsion inversion method based on the physical constraint neural network as claimed in any one of claims 1 to 7 is applied, and the icing torsion inversion system is characterized by comprising: The data acquisition module is used for acquiring torsion angle data, environment parameter data and historical icing thickness data of the power transmission wire; The physical model construction module is used for establishing a physical constraint model between the torsion angle and the ice coating thickness based on the wire mechanics principle; The neural network construction module is used for constructing a first neural network based on the physical constraint model and defining a composite loss function comprising a data driving loss function and a physical constraint loss function; The network training module is used for training the first neural network by utilizing the historical icing thickness data, optimizing network parameters through a back propagation algorithm, and minimizing the composite loss function; and the icing thickness inversion module is used for inputting the torsion angle data and the environmental parameter data into the trained first neural network for prediction to obtain the icing thickness predicted value of the current wire.
9. An electronic device comprising a memory and a processor, wherein the memory is configured to store computer-executable instructions, and the processor, when executing the computer-executable instructions, performs the steps of the method for ice-coating torsion inversion based on a physically constrained neural network as set forth in any one of claims 1 to 7.
10. A computer-readable storage medium having stored thereon computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of an icing torsion inversion method based on a physically constrained neural network as claimed in any one of claims 1 to 7.

Description

Icing torsion inversion method, system, equipment and medium based on physical constraint neural network Technical Field The invention relates to the technical field of power system monitoring, in particular to an icing torsion inversion method, system, equipment and medium based on a physical constraint neural network. Background Icing of a power transmission line is one of important disasters threatening the safe operation of a power grid. When the ice on the surface of the power transmission wire reaches a certain thickness, the power transmission wire can bear excessive load, so that serious accidents such as wire breakage, tower falling and the like are caused, and serious threat is formed to the safe operation of the power grid. Therefore, accurate monitoring of the icing thickness of the power transmission line has important significance for guaranteeing the safety of the power grid. The traditional icing monitoring method mainly relies on means such as manual inspection, image recognition and stress strain measurement, but with the improvement of monitoring requirements, the actual requirements of large-scale on-line monitoring of the power transmission line are difficult to meet by the traditional icing monitoring method. With the rapid development of artificial intelligence technology, the machine learning method provides a new technical path for icing monitoring. However, a machine learning model driven by pure data often needs a large amount of high-quality labeling data for training, and effective labeling data in a power transmission line icing monitoring scene is usually rare. The pure data driving model lacks physical understanding of the mechanical behavior of the transmission conductor, and the prediction result may violate basic physical rules, so that the generalization capability of the model is insufficient, especially in extreme weather conditions or in a novel conductor application scene. The physical information neural network is an emerging modeling paradigm in recent years, and the core idea is to embed a mathematical model describing a physical rule into the training process of the neural network in the form of constraint terms. According to the method, the neural network can be guided to learn the internal mode conforming to the physical rule under the condition of limited data, and the generalization capability and the interpretation of the model are improved. Therefore, aiming at the problems existing in the prior art, it is highly desirable to provide an icing torsion inversion method based on a physical constraint neural network. By embedding the wire mechanical physical model into the neural network training process, the accurate inversion from the torsion angle to the icing thickness is realized. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides an icing torsion inversion method, system, equipment and medium based on a physical constraint neural network, which solve the problems that the existing icing monitoring technology has poor instantaneity, insufficient precision, high cost or depends on a large amount of labeling data, and the effective fusion of the mechanical rules of wires is lacking, so that the actual requirement of online reliable icing monitoring of a large-scale power transmission line is difficult to meet. In order to solve the technical problems, the invention provides the following technical scheme: In a first aspect, the invention provides an icing torsion inversion method based on a physical constraint neural network, which comprises the following steps: Collecting torsion angle data, environmental parameter data and historical icing thickness data of a power transmission wire; establishing a physical constraint model between a torsion angle and the ice coating thickness based on a wire mechanics principle; constructing a first neural network based on the physical constraint model, and defining a composite loss function comprising a data-driven loss function and a physical constraint loss function; Training the first neural network by using the historical icing thickness data, and optimizing network parameters by a back propagation algorithm to minimize the composite loss function; And inputting the torsion angle data and the environmental parameter data into a trained first neural network for prediction to obtain the icing thickness predicted value of the current wire. As an optimal scheme of the icing torsion inversion method based on the physical constraint neural network, the method for constructing the first neural network based on the physical constraint model comprises the following steps: Determining the number of input layer nodes, the corresponding torsion angle and the environmental parameters of the first neural network; configuring a hidden layer structure of the first neural network by adopting a multi-layer full-connection netw