Research On Intelligent Evaluation Methods Of Dynamic Frequency Response Characteristic Metrics Under Credible Contingencies

Since the massive integration of renewable generation and the construction of HVDC transmission projects,the system inertia level and the primary frequency regulation capability have been greatly reduced.As a result,the risk of frequency instability under the large-capacity active disturbances has increased dramatically.At present,the full model time-domain simulation method is widely used to simulate the dynamic frequency spatial-temporal distribution characteristics after expected disturbances(i.e.,large-capacity units outage,DC blocking)and obtained dynamic frequency response characteristic metrics,such as extreme frequency,quasi-steady-state frequency,and maximum frequency change rate.However,due to the inherent problems of model complexity and excessive computational time,the full model time-domain simulation method is only suitable for offline analysis and difficult to be applied to online evalution of dynamic frequency response characteristics metrics under disturbances.For the power system with high proportion of renewable energy,fluctuating power such as wind power and photovoltaics have increased the diversity and complexity of operation modes.In the case of “combinatorial explosion” in view of multiple uncertainties(stochastic wind/solar generation,load fluctuation,contingencies,etc.),time-domain simulation is difficult to meet the needs of fast dynamic frequency response characteristics metrics evaluation.Consequently,there is an urgent need for a more convenient,intelligent and online application method to evaluate the dynamic frequency response characteristics metrics under credible contingencies so that dispatchers and planners can quickly and accurately grasp the dynamic frequency response characteristics of power system,and then effectively formulate preventive control measures to ensure that the frequency after the disturbances can be tolerated,and maintain the safe and stable operation of the system.The main research contents and achievements in this paper include:(1)Due to the analytical ability of typical features of the input data is insufficient and it is difficult to map the input-output relationship under complex conditions.Also,the problems of over-fitting or under-fitting are difficult to resolve and the generalization ability is poor,an intelligent evaluation method of dynamic frequency response characteristics metrics under credible contingencies based on stack denoising auto-encoder(SDAE)is proposed.The method is divided into two major functional modules: offline training and online evaluation.In offline training,SDAE uses the“pre-training,fine-tuning” method to train network parameters.In the pre-training process,the Dropout method was used to improve algorithm generalization ability and prevent over-fitting.Then,the root mean square back propagation(RMSprop)optimization was deployed to fine-tune network parameters so as to reduce the possibility of falling into local optimums.The effectiveness and superiority of the SDAE in the intelligent evaluation of dynamic frequency response characteristics metrics under credible contingencies are verified by using the IEEE RTS-79 and a provincial power grid as an example.(2)In order to solve the problem of long off-line training time and low prediction accuracy under a small number of training samples,an intelligent evaluation method of dynamic frequency response characteristics metrics under credible contingencies based on XGBoost(extreme gradient boosting)is proposed.And we use Bayesian optimization method to realize automatic optimization of XGBoost hyperparameters.A provincial power grid is selected for case study,and the comparison with SDAE verified the superiority and rapidity of XGBoost in the intelligent evaluation of dynamic frequency response characteristics metrics under credible contingencies.(3)To solve the problem that the pure data-driven method is difficult to accurately predict the dynamic frequency response characteristics metrics of small probability edge samples when the training samples are extremely small,an integrated data-driven-model-driven intelligent evaluation method is proposed.The strong causal relationship between electrical information is preserved by the aggregate system frequency response model,which ensures that a reliable solution can be obtained in a small probability disturbance.The optimal identification of the dynamic damping factor corresponding to the nadir of the frequency and quasi steady state frequency is realized by data-driven method,which further improves the accuracy of dynamic frequency response characteristics metrics prediction based on aggregate system frequency response model.A provincial power grid is selected to carry out case sudy,and compared with SDAE and XGBoost to verify the superiority of integrated data-driven-model-driven in the intelligent evaluation of dynamic frequency response characteristics metrics under credible contingencies.