Research On Hierarchical Artificial Intelligence For Online Total Transfer Capability Calculation And Regulation

With the increasing scale of interconnected power grids and the growing uncertainties,operating condition of power transfer inter-corridors shows the trend of fierce variation near security boundary.To maintain security of power systems,it is vital to monitor condition of inter-corridors in real-time.Dispatchers are used to take total transfer capability(TTC)to measure inter-corridors’ security situation,further,to proactively control TTC beyond power transfer to guarantee safe operation.However,the inclusion of dynamic security verifications among TTC determinations causes a great computational burden.Besides,manifold transient behaviors introducing by dynamic machines,as well as coupling patterns between intercorridors,result in unavailability of model-based TTC computation.Artificial intelligence(AI)and data-driven technique(DDT),which in possession of model-free feature,can offline learn power system patterns and online assess targets via simple matrix operation.AI allows us to prevent sophisticated physical modelling,and is a potent alternative to realize online TTC calculation and regulation.There are two pivotal challenges of applying AI to address TTC-oriented problems.The one is that how to choose proper power system features to help AI capture TTC behavior.The other is the proper synergy of “black-box” AI with decision-making program.This dissertation aims at designing data-driven scheme for online TTC calculation and control.On the basis of AI,a TTC-oriented hierarchical framework is constructed,which encompasses real-time calculation,preventive control and temporal operation decision from the bottom up.The salient contributions are listed below:(1)In order to enhance security monitoring capacity for wind power exporting corridors,data-driven/AI techniques are carried out to establish TTC calculation rules.It is demonstrated that,the proposed data-driven methods enable second level TTC calculations with satisfied accuracy,and deep learning outperforms existing linear regression methods over 38% regarding TTC estimation accuracy.(2)A surrogate-assisted framework to efficient infer security of corridors is proposed: consider it that AI methods markedly reduce the complexity and dimensionality of physics-based TTC model and provide parameterized/explicit TTC rules against critical variables,AI-based TTC approximate is utilized to substitute the most complicated and time-consuming part inside decision model.In this manner,security inference can be done in real-time,so that online TTC regulation is attainable.(3)Imperative error of AI techniques may cause erroneous inference for system security,further leading to risky regulations.To address this problem,prediction interval technique is introduced to enable a confidence-aware deep learning surrogate,which can be further incorporated into TTC regulation for conservative preventive control.The proposed strategy enables improvement on reliability and dependability of AI-guided security control.(4)A real-time TTC-oriented preventive control strategy is proposed through deep reinforcement learning.To circumvent computational burden from slow interaction between AI agent and TTC control model,surrogate-assisted method is employed to simplify the interactive environment.Deep learning for TTC calculation underlies this strategy,forming seamless cooperation for multiclass AIs.The proposed hybrid AI-based strategy is demonstrated to be able to provide TTC control actions in seconds,and is strongly expected to be insensitive to grid size.(5)A rolling-horizon operational planning constrained with dynamic TTC is studied.For temporal considerations(e.g.,wind power,energy storage,unit ramp,etc.),the former surrogate assisted TTC regulation under single time snapshot is extended to multi-period.Further to fast solve this high-dimensional problem,a gradient-based decision algorithm is designed.The algorithm also allows us to derive Jacobian and Hessian matrices of precise deep learning responses vs.decision variables.Thus,the temporal operational planning with dynamic TTC constraints can meet both high accuracy and efficiency.Numerical studies show that the proposed method outperforms model-based methods over 20 times in decision speed.At the end,we would also like to note that,the proposed AI-based methods in this dissertation are of generality and scalability,can be readily extended to other problems with unacceptable time-/resource-intensive simulations/experiments.