Abstract:

China is one of the world’s most earthquake-prone countries, and earthquake prediction is critical for mitigating disaster impacts. To detect precursors, China has built the world’s largest seismic precursor observation network, with geoelectric field observation as a key method. However, geoelectric field observations are affected by multiple factors, such as interference from high-voltage direct current(HVDC)transmission and subway operations. HVDC’s impact is growing with recent infrastructure expansion. Unbalanced operation of HVDC parallel lines severely distorts geoelectric field data, causing large-amplitude, wide-ranging interference that impairs data usability for earthquake forecasting studies. Disturbance patterns vary across regions and lines, with spikes, steps, and combinations as typical HVDC-induced anomalies. Automated recognition remains underdeveloped; current practices rely on manual monitoring to identify affected stations and periods, consuming substantial labor and time.
To enhance the efficiency of identifying disturbed data while ensuring accuracy, this study proposes a Bi-LSTM-based interference recognition method for the automatic identification of HVDC interference types and periods in geoelectric field data, drawing on existing research on HVDC-induced geoelectric interference mechanisms and typical patterns. The method aims to reduce reliance on expert knowledge and provide a reliable basis for subsequent data quality control and interference correction. Geoelectric field data present regular diurnal variations and distinct temporal characteristics, while HVDC interference signals feature long durations and complex morphological evolution, making unidirectional temporal modelling insufficient to fully capture their dynamic features. Compared with other deep learning methods, Bi-LSTM can simultaneously utilize forward and backward temporal information to capture dependencies from past and future time points, achieving higher recognition accuracy and robustness than unidirectional LSTM or other time-series models, thus effectively capturing implicit temporal features in geoelectric data and realizing efficient and accurate identification of disturbed anomalies to meet technical requirements. To ensure sufficient samples for model training, simulated interference is superimposed on normal geoelectric data to generate three main types of disturbed data. Samples are divided into training, validation and test sets: the training set is used to train the model’s internal neural parameters, the validation set for hyperparameter adjustment, and the test set to evaluate the model’s generalisation ability.
The trained model achieved 96.6% accuracy and precision on the test set. To verify its generalization ability and reliability in real scenarios, the Bi-LSTM model was applied to actual disturbed geoelectric field data. Results show its automatic interference labels are highly consistent with manual annotations. These indicate the Bi-LSTM method can model temporal dependencies in geoelectric data and recognize nonlinear interference patterns to a certain extent. Compared with conventional approaches, it performs better in interference event localization and classification, showing potential applicability for HVDC-induced disturbances.
The results of this study can greatly improve the efficiency of interference recognition, laying the foundation for the subsequent processing and effective use of interference data, while also providing new ideas and technical references for recognizing other factors that interfere with geoelectric field observations, such as power supply interference affecting geoelectric field measurements in co-located station resistivity observations.

Key words: HVDC, LSTM, Geoelectric field observations, interference recognition

摘要：

近年来, 随着中国高压直流输电线路的不断建设, 高压直流输电线路对地电场观测数据的影响日趋严重。现有识别高压直流输电对地电场干扰事件的方式主要依靠人工, 效率较低, 深度学习方法能够显著提高对干扰事件的识别效率。其中, 基于深度学习的双向长短期神经网络(Bidirectional Long Short-Term Memory, Bi-LSTM)具有对时序数据强大的建模和特征提取能力, 能够捕捉时序数据前、 后2个方向的长短期信息依赖关系。因此, 为了高效准确地对地电场观测数据中受高压直流输电干扰的时段进行识别, 文中采用Bi-LSTM模型实现对高压直流输电对地电场观测干扰时段和干扰事件类别的识别。最终模型在测试集上达到了较高的精度(96.6%)和较高的准确率(96.6%), 证实了模型的泛化性和实用性。该模型验证了Bi-LSTM在地电场受高压直流输电干扰识别方面的应用效果, 为后一步对高压直流干扰数据处理提供了良好基础, 也为地电场观测数据受其他干扰因素的识别和分类提供了一种新的思路。

关键词: 高压直流输电, 长短期记忆神经网络, 地电场观测数据, 干扰识别