Research On Short-term Load Forecasting Based On Distributed Graph Computing

Load forecasting is the basic work of power system planning,operation and scheduling.It can be used for the schedule,the optimal combination of units and the optimal power flow to ensure the safety and economic operation of the power grid.Meantime,load forecasting results also provide the basis for the annual maintenance plan,investment planning,medium term operation and electricity marketing.With the expansion of the power grid,the data produced by the power system has reached the scale of large data.It provides the data basis and practical basis for accurate load forecasting based on large data technology.The power load is easily affected by the weather conditions,holidays and other factors in the region.The dissertation studies the load forecasting from the perspective of multi factor correlation.The dissertation analyzes the daily type,temperature and humidity as factors,and gives the load change trend.A DBN model is proposed with theoretical derivation,which is based on the factors of week,humidity and temperature.Finally,Taking the transformers as an example,the accurate load forecasting is realized,and the feasibility of the proposed algorithm is verified.Load forecasting based on distributed computing of graph structure is proposed.This dissertation designs a distributed computing platform and study the specific implementation of the algorithm.For empty data,abnormal data and duplicate data,the MapReduce parallel computing method is used to modify the data,and data quality check is conducted.The state division of load,week,temperature and humidity data are realized by the method of equidistant partition.The K-Means clustering is used to replace the certain transformer with the cluster in order to enlarge the sample size.Load,week,temperature and humidity data are loaded into the graph database,and the graph structure is established.Finally,all kinds of probabilities are calculated in parallel with the traditional matrix traversal and graph edge path retrieval.According to the probability graph algorithm,the predicted values of all the transformers in the region are given.The experimental results show that the efficiency of graph computing is significantly higher than that of matrix operation and the feasibility of distributed graph computing method is verified.