Cluster-based Distribution Network Ultra-short-term Load Forecasting And Economic Optimal Dispatch

This topic is one of the main contents of the science and technology project “Research on Key Technologies of Planning and Control of Wide Area Distributed Power Based on Big Data Analysis”,supported by Shanxi Electric Power Company.In recent years,with the full development of the construction of user power consumption information collection systems,the data generated in the operation,monitoring,and management of power grids has grown in a geometrical manner,with a large number of features,how to process data and obtain effective information is a new opportunity and challenge.Clustering as an important means,has shown certain advantages in mining analysis.It can effectively extract the potential model information of electricity consumption curve and support the personalized and differentiated demand of electricity service,and will be widely used in the future power system field.Load characteristics analysis and recent short-term load forecasting are the decision-making basis for grid companies to make scheduling and distribution network planning.The economic dispatching optimization decision of the distribution network is a key means to maintain the efficient and safe operation of the distribution network,and is also the core technology for the distribution network to actively manage the distributed power supply.In response to the above-mentioned issues,based on the overall consideration of the development trend of short-term load forecasting technology,this thesis starts with the analysis of electricity user behavior analysis,and an integrated model for short-term load forecasting based on cluster analysis is constructed to support the dynamic adjustment and accuracy of short-term forecast results,and to provide an effective basis for economic optimal dispatching of distribution networks.The main research content is as follows:An in-depth analysis of the input variable characteristics of the short-term load forecasting model.On the basis of studying the daily,weekly,and monthly load characteristics of the power grid,the Spearman rank coefficient method was used to analyze the relationship between load forecasting and its influence factors,and the most significant daily maximum temperature and humidity were taken as the meteorological input factors of the model.Preprocessing loads,meteorological data samples,identify and correct missing and abnormal points,normalize load data,and quantify temperature data.Aiming at the feature variable selection mode in the data clustering process,a load characterization dimension reduction optimization strategy based on the improved Relief method is designed,by selecting feature variable sets that are closely related to the power user behavior pattern and have higher independence,the cumbersomeness of the clustering process is reduced and the clustering effect is improved.A Bayesian regularized SOM clustering model was established based on the optimal characteristics.The model was studied and clustered under unsupervised classification.After the UCI data set was used for validation,the actual load data sample of the power grid was analyzed.The example proves that the method proposed in this thesis can effectively identify users of four different power consumption modes: two-shift enterprise,elderly family,business and college load.Analyzing user load curves is of great importance for achieving high-precision prediction and ensuring system stability.Based on the concept of distributed computing,the prediction process is disassembled to various types of users.Through classification prediction of users with different power consumption modes,user resources are fully utilized to optimize the user's future load forecasting effect and effectively enhance model fitting performance.Aiming at the shortcomings of the conventional extreme learning machine for solving the cumbersome and batch training,an improved nuclear limit learning machine model is constructed.The complex inverse change is simplified by the cholesky decomposition,and the incremental algorithm is constructed to enhance the update of the model.The model is applied to the actual special variable load forecasting,and comparing the results before and after the improvement and the traditional Elman neural network prediction results,the improved KELM has higher precision and faster speed.Considering the shortcomings of the traditional "subnet accumulation",a full-network load model was constructed based on the sub-network load stability and proportional coefficient prediction,which reduces the error MSE by 0.0547,significantly improves the model generalization ability and provides a basis for making scientific scheduling decisions.An opportunity distribution planning optimization model for economic dispatching of regional distribution networks was constructed.Taking into account distribution network reliability constraints,and aiming at maximizing the operating economic efficiency of a complete dispatch cycle of the distribution network,using controllable distributed energy as a control method,the impact of real-time purchase and sale price adjustment on operating costs is considered,the model is solved using a harmony search algorithm that combines Monte Carlo simulations.Finally,taking the actual system of a certain region as an example,the universal applicability of the model and the solution algorithm is verified,and the optimal scheduling strategy for operating costs is obtained.Analyze the conservativeness of different confidence level control decisions and achieve the coordination between economy and reliability.The project has passed the acceptance meeting of the science and technology project of Shanxi Electric Power Company of the State Grid,which provides a theoretical basis for the safe and economic operation of the power grid system.