Research On Power System Static Voltage Stability Assessment And Real-time Economic Dispatch Based On Machine Learning

Peaking carbon emissions and achieving carbon neutrality have become the common goals all over the world and will be the energy trend in the future.The increase of the proportion of renewable energy generation is an important way and determining factor to promote the carbon emission reduction.The rapid growth of renewable energy generation not only brings historic opportunity but also new issues and challenges for electric power system.First,the optimization and upgrading of energy structure has raised new requirements on the accurate prediction of renewable energy generation.However,the inherent intermittency,randomness and volatility of primary energy,such as wind or solar energy,greatly increase the difficulty of prediction.Second,the large-scale energy storage technology is not yet mature.To enhance the large-scale power grid construction and encourage the application of interactive energy-using equipment are necessary means for promoting the renewable energy development.However,the complexity of the structure and changes in load characteristic make the power system static voltage stability issue more prominent.The power system economic dispatch needs to maintain the voltage stability for safe and steady power system operation,before achieving the economic goal.How to achieve the real-time economic dispatch that can cope with the rapid development of renewable energy and voltage stability has become an important research topic in power system.Although the high proportion of renewable energy connected brings the challenge to the power system,the new machine learning technologies such as deep learning and ensemble learning,provide new methods and new ideas for solving practical engineering problems in power system.Against the above background,this thesis firstly researches the centralized photovoltaic power ultra short-term forecasting method based on attention mechanism.Then it uses the forecast results to construct the power balance equation in real time economic dispatch optimization model with load forecasting data.Secondly,this thesis researches the power system static voltage stability margin assessment based on machine learning.The proposed assessment model is used to construct the static voltage stability margin reliable assessment method based on ensemble learning.The extracted assessment rules that are obtained by offline analysis form the static voltage stability constrained conditions in economic dispatch optimization model.Finally,this thesis studies the real time power system economic dispatch based on deep learning and treats it as a learning problem instead of an optimization problem.The main contents are as follows:(1)This thesis proposes an accurate prediction method of photovoltaic power based on attention mechanism for centralized photovoltaic power ultra short-term prediction problem.The method makes the power balance equation in economic dispatch optimization model considering the photovoltaic power and reduces the uncertainty effectively.First,Spearman correlation coefficient method and grey relational analysis are used to analyze the photovoltaic power characteristics such as key meteorological factors,temporal correlation and spatial relational.Second,the structure of Long Short-term Memory(LSTM)network is designed in an adaptive manner based on the results of characteristics analysis.The photovoltaic power forecasting model is then constructed by the adaptive LSTM network.Finally,the photovoltaic power sequence is further classified as spatial-relational time series data.The attention mechanism is introduced into photovoltaic power characteristic analysis and the ultra short-term photovoltaic power accurate forecasting model is then built based on Multi-level Nesting Spatial Temporal-Attention Network(MNST-AN).This accurate forecasting method changes the inherent process of characteristic analysis before using machine learning to solve photovoltaic power prediction problem and avoids the influence of human factor.It can promote the forecasting accuracy practically.The feasibility and accuracy of the photovoltaic power ultra short-term forecasting model are verified by actual data obtained from a centralized photovoltaic plant in North China.(2)This thesis proposes a static voltage stability margin assessment method based on decision tree for power system static voltage stability issue to construct the static voltage stability constrained condition in economic dispatch optimization model.First,the system status is divided into normal state,alert state and emergent state by using P-V curves analysis.The training samples are then obtained from each state.Second,the feature variables are pre-selected by using participation factors analysis on the physical meaning.In order to solve the miss alert issue,feature variables are further selected from the capacity of data classification by using the proposed Relief-F-P feature selection algorithm,which is based on Relief-F algorithm and considers penalty factor.Finally,the decision tree for voltage stability margin assessment is constructed by C4.5 algorithm.The effectiveness of the proposed method is validated on actual data obtained from a load center in South China.(3)Furthermore,this thesis proposes a static voltage stability margin reliable assessment method based on random forest to further improve the accuracy.First,a training sample subsets construction method is proposed by extracting P-V curves randomly,which can balance the diversity and learning ability of basic decision tree.Second,a split-attribute selection method is proposed based on information gain ratio to improve the controllability of introduced disturbance.Based on the generated training sample subsets and split attributes,the C4.5 algorithm is used to construct the basic decision tree.The majority voting strategy is then used to construct the random forest learning model for static voltage stability margin assessment.Finally,a ruled score calculation method that integrated coverage and reliability is proposed to extract key rules.Case study on a real system in South China demonstrate that the proposed method can assess the system voltage stability effectively.(4)This thesis proposes a new idea for solving power system real-time economic dispatch based on deep learning.First,the power balance equation and static voltage stability margin assessment rules in economic dispatch optimization model are constructed using photovoltaic power forecasting results and the extracted key assessment rules.Second,the training targets of learning model are built by the optimization model based on massive historical dispatch data,which is guided by perfect dispatch theory.Then,the differentiated training sample sets are constructed by hierarchical cluster analysis and matrix correlation analysis.Based on the differentiated training sample sets,the structure of the Gated Recurrent Unit(GRU)is designed adaptively and the real-time economic dispatch learning model based on adaptive GRU is proposed.Once the learning models are determined,the dispatch schedules can be obtained by inputting corresponding forecasting data.The feasibility and the effectiveness of the proposed model is verified by IEEE-39 test system.The proposed model in large-scale power grid is further verified in IEEE-118 test system.The studies can enrich the research theory of the power system dispatching operation and improve the intelligence and automation.It can also accelerate the carbon emission reduction for electric power industry and promote the implementation of peaking carbon emissions and carbon neutrality.