基于LSTM-CNN-Attention模型的电力设施非周期巡视决策方法

Non-periodic inspection decision method of power facility based on LSTM-CNN-Attention model

随着电力系统规模的日益增大,电网面临不确定性故障的危险,会影响人们的日常生活,甚至可导致重大安全事故。因此,提前预测电力设施的运行状态并作出巡视修检决策非常重要。但常用的决策方法(如支持向量机(SVM)模型等)在这些实际应用场景中存在准确度不高、召回率低的问题。针对这一问题,提出一种结合长短期记忆(LSTM)、卷积神经网络(CNN)和注意力(Attention)机制的电力设施非周期巡视决策方法LSTM-CNN-Attention,将数据经过极限梯度提升(XGBoost)特征选择和归一化处理后输入该决策模型,利用注意力机制对经过LSTM和CNN层提取的包含时间和空间的信息作加权处理,区分信息的重要程度,以在输出预测结果时能够更关注那些对结果影响最大的信息,确保在预测过程中更重要的信息能够得到更大的关注和贡献,以提高预测结果的准确性和可靠性。通过在电力设施运行数据集上进行对比实验,验证了LSTM-CNN-Attention的准确率、精确率、召回率和F1-score性能评估指标优于CNN-LSTM、XGBoost、CNN、随机森林、SVM和逻辑回归模型的学习算法。

With the increasing scale of the power system,the power grid is facing the risk of uncertain faults that can affect people’s daily life and can even lead to major accidents.Therefore,it is very important to predict the operating status of power facilities in advance and make inspection and repair decisions.However,commonly used decision-making methods,such as SVM(Support Vector Machine)model,have problems of low accuracy,low recall,in these practical application scenarios.Aming at the issue,a non-periodic inspection decision-making method LSTM-CNN-Attention was proposed for the non-periodic inspection of power facility by combining Long Short-Term Memory(LSTM),Convolutional Neural Network(CNN)and Attention mechanism.The data was input into the decision model after XGBoost(eXtreme Gradient Boosting)feature selection and normalization,and attention mechanism was used to weignt the temporal and spatial information extracted by the LSTM and CNN layers,so as to differentitate the importance of the information that was most influential to the result,thus ensuring that the more important inforamtion could receive greater attention and contribution during the prediction process,and the accruracy and reliability of the prediction results could be improved.Through comparative experiments on power facility operation datasets,it is verified that the proposed method is superior to the learning algorithms of CNN-LSTM,XGBoost,CNN,random forest,SVM and logistic regression models in accuracy,precision,recall and F1 score.