基于GRA-CEEMDAN-BiLSTM的变压器油中溶解气体浓度预测

Prediction of Dissolved Gas Concentration in Transformer Oil Based on GRA-CEEMDAN-BiLSTM

基于自适应噪声完全集成经验模态分解与双向长短期记忆神经网络的变压器油中溶解气体浓度预测模型,本文中作者提出一种考虑待预测状态参量与影响因素相关性。首先,利用灰色关联分析法确定与待预测状态参量关联度较高的因素,作为关键预测输入参量;然后,运用自适应噪声完全集成经验模态分解将待预测参量序列分解为不同尺度相对平稳的子序列分量,再将关键输入参量序列与分解得到的子序列组成输入矩阵分别输入各双向长短期记忆神经网络模型,构建各子序列的预测模型;最后,将各组分量预测结果进行叠加重构,得到最终的气体浓度预测值。通过实例验证了所提模型的可靠性和有效性。

A complete ensemble empirical mode decompositionwith adaptive noiseand bidirectional long short-term memory-based predicting model of dissolved gas concentration in transformer oil is proposed. The model considers the influence of transformer oil temperature, load, and ambient temperature on the gas concentration in the oil. First,the grey relation analysis method is used to determine the factors that have a higher correlation with the state parametersto be predicted, and extract them as key forecast input parameters;Then, the complete ensemble empirical mode decomposition with adaptive noiseisused to decompose the parameter sequence to be predicted into a group of relatively stable sub-sequence components, and then combining the correlation relationship to form the input matrix of the key input parameter sequence and the subsequence obtained by decomposition and input into each bidirectional long short-term memory model respectively to construct the sub-sequence prediction model;Finally,each group of prediction components is superimposed and reconstructed to obtain the final gas concentration prediction value. The reliability and effectiveness of the proposed model are verified by an example.