基于融合算法的支撑电容寿命建模

Lifetime Modeling of Support Capacitor Based on Fusion Algorithm

摘要:

为弥补单一算法在薄膜电容器寿命预测中的不足、提高预测精度，针对电容量损失时间序列样本量小、测量点少的特点，采用单步（Single-step Forecast Strategy，SFS）、多产出（Multiple Output Strategy，MOS）以及递归多步（Recursive Multi-step Forecast Strategy，RMFS）预测策略对其进行综合描述，然后通过分析深度置信网络（Deep Belief Network，DBN）、支持向量回归（Support Vector Regression，SVR）和BP神经网络的性质，提出融合人工智能学习算法的三种时间序列预测模型对电容器寿命进行估算:基于误差补偿的DBN?SVR?RMFS用于早中期寿命估算，DBN?BP?MOS用于晚期趋势预测，SVR?SFS用于结合检修计划的阶段性预测。结果表明，三类模型在不同工况下的预测误差均在允许范围内，提高了预测精度与鲁棒性，满足电容器全寿命周期状态评估与更换决策的需求，验证了模型的有效性。

Abstract:

To compensate for the shortcomings of single algorithms in the lifetime prediction of metallized film capacitors and to improve prediction accuracy, given the characteristics of small sample size and few measurement points in the capacitance loss time series, the single-step forecast strategy (SFS), multiple output strategy (MOS), and recursive multi-step forecast strategy (RMFS) are adopted for comprehensive description. Subsequently, by analyzing the properties of the deep belief network (DBN), support vector regression (SVR), and BP neural network, three time series prediction models integrating artificial intelligence learning algorithms are proposed to estimate the capacitor lifetime: the error compensation-based DBN-SVR-RMFS for early and mid-term lifetime estimation, the DBN-BP-MOS for late-term trend prediction, and the SVR-SFS for periodic prediction combined with maintenance schedules. The results demonstrate that the prediction errors of the three types of models under different operating conditions are within acceptable ranges, indicating improved prediction accuracy and robustness, satisfying the requirements for full-lifecycle condition assessment and replacement decision-making of capacitors, and validating the effectiveness of the models.