A Novel Method for Predicting Smart Grid Stability Via DNN and Hybrid Ensemble Strategy

Smart grid stability prediction (SGSP) is one of the key issues in the energy field, due to its relation to national security and the interests of the people. Predicting whether the smart grid is in an unstable state or not can help remind grid managers to take preventive measures in advance to reduce the harm caused by grid instability. For this issue, we propose a novel SGSP method named SS-DNN based on deep neural network (DNN) and hybrid ensemble strategy, i.e., snapshot stacking (SS) ensemble strategy. Specifically, the DNN has a four-layer structure with balanced cross entropy loss for solving class imbalance problem, and the training process of SS-DNN for SGSP involves two stages. The first stage is to train multiple DNN models with diversity. Under the condition of the same DNN structure, these DNN models are trained by cosine annealing learning strategy from snapshot ensemble algorithm for model diversity. The second stage is to combine these DNN models with support vector machine (SVM)-based meta learner, inspired by stacking ensemble algorithm. Simulation results show that the single DNN-based SGSP method has better performance than other machine learning methods, and its accuracy reach up to 98.715%. Moreover, the SS-DNN for SGSP can reach up to 99.255%, which outperforms the single DNN and exceeds other ensemble methods.