Study On Power Systems Transient Stability Assessment Based On Machine Learning Method

An electrical power system is often subject to various big disturbances during operation, which can cause transient instability, especially line-to-ground fault and short circuit fault. The damage of transient instability stands as the one of the main factors resulting in catastrophic accident of power grid. Quick and accurate transient stability assessment methods are of great importance to guarantee the power grid's safe and stable operation. On the other hand, electrical power system planning also requires the analysis of overall transient stability level.As artificial intelligence is developing rapidly, researchers begin to pay attention to transient stability assessment methods of power system based on machine learning. From the point of machine learning, power system transient stability assessment is a kind of Pattern Recognition. Namely some mapping relation exists between transient stability and some features describing system operation status. Through offline numerical analysis, samples are obtained that can sufficiently reflect this kind of mapping characteristic. Meanwhile, we can gain the function relation for unknown mapping by studying the samples. Once the function relation is got, analyzing result of system transient stability is achieved under new operation status by the mapping relationship between features and stability. Therefore, the method based on machine learning stands as the most promising approach to study power system transient stability assessment.In this PhD thesis, we studied a number of issues linked to power system's transient stability assessment based on machine learning method, such as feature selection, combined assessment model, severely disturbed machine's input feature, online learning, probabilistic assessment and etc.(1) According to redundancy among input features and independency between some features and power system transient stability, a dual stage feature selection method based on SVM was put forward. First of all, the original features are sorted using recursive feature element method and removed of those behind and the features at the front are chosen to form feature subset. Then the Wrapper method was adopted using support vector machine as classifier and near-optimal feature subset was obtained by best first search. The emulation result of New England 39-bus test system and IEEE 50-generator test system show the feature selection method can help to get near-optimal feature subset.(2) In order to improve the performance of power system transient evaluating model, three transient stability assessment models based on Meta learning were propoed in the thesis, including Stacking frame which ensembled decision tree, SVM and naive Bayesian, AdaBoost method and RandomForest method to combine decision trees. The emulation result reveals the above three methods are all superior to single models. The AdaBoost and RandomForest method bring the best assessment performance.(3) The power system transient stability assessment based on severely disturbed generator attributes and machine learning method was studied. Candidate machines were detected by synthesizing three severely disturbed machine recognition methods. One is machine corresponding to the max value of relative kinetic energy at the initial moment of a fault, another is that corresponding to initial acceleration power of unitary and the other corresponding to the most severely disturbed generator's acceleration speed when losing stability. From the three methods, the 42-dimention features were got based on the three severely disturbed machines. The emulation on New England 39-bus test system and IEEE 50-generator test system shows the severely disturbed features can effectively represent system dynamics.(4) An online learning assessment method of power system transient stability supporting SVM incremental learning was brought forward. Through structuring recursive method, the new data were added, but the original data were kept as KKT. The emulation reveals incremental learning can help to reduce learning time dramatically and is a promising online learning method in transient stability assessment.(5) Monte Carlo-support vector machine transient stability probabilistic assessment method was studied in this thesis. Emulation samples were built by use of non-sequential Monte Carlo and support vector machine was used to accelerate transient stability assessment. From the result, the above method can reduce assessment time as well as keep precision.(6) The non-sequential Monte Carlo does not take quasi-random sequences into consideration, so Markov Chain Monte Carlo Method was introduced into power system transient stability assessment. New model and method which were based on MCMC for probabilistic transient stability assessment were put forward. The emulation result shows that the method is quicker and brings better stability than traditional Monte Carlo approach.