何秀思,阮方鸣,徐愷,等. 基于Attention-LSTM-XGBoost的电极移动速度影响放电参数预测分析[J]. 电波科学学报,2024,39(2):1-9. DOI: 10.12265/j.cjors.2023154
      引用本文: 何秀思,阮方鸣,徐愷,等. 基于Attention-LSTM-XGBoost的电极移动速度影响放电参数预测分析[J]. 电波科学学报,2024,39(2):1-9. DOI: 10.12265/j.cjors.2023154
      HE X S, RUAN F M, XU K, et al. Analysis of discharge parameter prediction affected by electrode movement speed based on Attention-LSTM-XGBoost[J]. Chinese journal of radio science,2024,39(2):1-9. (in Chinese). DOI: 10.12265/j.cjors.2023154
      Citation: HE X S, RUAN F M, XU K, et al. Analysis of discharge parameter prediction affected by electrode movement speed based on Attention-LSTM-XGBoost[J]. Chinese journal of radio science,2024,39(2):1-9. (in Chinese). DOI: 10.12265/j.cjors.2023154

      基于Attention-LSTM-XGBoost的电极移动速度影响放电参数预测分析

      Analysis of discharge parameter prediction affected by electrode movement speed based on Attention-LSTM-XGBoost

      • 摘要: 基于具有Attention机制的长短期记忆(attention long short-term memory,Attention-LSTM)神经网络模型,设计了一种由Attention-LSTM神经网络与极端的梯度增强 (extreme gradient boosting,XGBoost )法共同组成的变权组合模型,用以分析预测静电放电过程中电极移动速度对放电参数造成的影响。该组合模型充分利用静电放电参数的时序特性,并采用Attention机制突出对放电参数预测起到关键作用的输入特征。首先基于由新型电极移动速度效应测试仪的实验结果提供的原始实验数据,采用分箱法对其进行预处理得到新的实验数据;然后将得到的新实验数据集作为两种模型的输入数据,分开训练Attention-LSTM模型和XGBoost模型,求解出各自模型的预测结果及误差;最后利用误差倒数法,重新计算出两种模型预测结果的占比权重,并根据计算的权重求解出最终预测结果。预测结果表明:与Attention-LSTM神经网络模型、XGBoost模型、Attention-LSTM-XGBoost定权组合模型相比,本文构建的Attention-LSTM-XGBoost变权组合模型,评估指标中的决定系数分别提升了5.22%、9.11%、3.13%。本文提出的变权组合模型在预测精度以及算法鲁棒性上均优于其他模型,有益于对小间隙静电放电参数变化趋势和规律的探寻。

         

        Abstract: Based on the long short-term memory neural network model with Attention mechanism (Attention-LSTM), a combined variable weight model consisting of Attention-LSTM neural network and Extreme Gradient Boosting (XGBoost) is used to analyze and predict the effect of electrode movement speed on the discharge parameters during electrostatic discharge. The combined model takes full advantage of the time-series characteristics of electrostatic discharge parameters and uses the Attention mechanism to highlight the input features that play a key role in the prediction of discharge parameters. Based on the original experimental data provided by the experimental results of the newly developed electrode velocity effect tester, the analytical data were preprocessed using the split-box method to obtain brand-new experimental data. Then the obtained new experimental dataset is used as the input data for the two models, and the Attention-LSTM model and XGBoost model are trained separately, and the prediction results and errors of the respective models are solved respectively. Afterwards, by using the error inverse method, we recalculated the proportional weights of the prediction results of the two models and solved the final prediction results according to the calculated weights. The control group models in the paper are: Attention-LSTM neural network model, XGBoost model, and the fixed-weight combination model. The prediction results show that compared with the basic Attention-LSTM neural network model, XGBoost model, and the fixed-weight combination model, the variable-weight combination model constructed in this paper improves the coefficient of determination in the evaluation index by 5.22%, 9.11%, and 3.13%. The combined model proposed in this paper outperforms other models in terms of prediction accuracy and algorithmic robustness, and is useful for exploring the trends and patterns of small gap electrostatic discharge parameters.

         

      /

      返回文章
      返回