Short-term Load Forecasting Based On Load Pattern Extraction

Short-term load forecasting is an indispensable part of ensuring grid safety and economic operation.Accurate forecasting results are conducive to improving the utilization rate of power generation equipment and the effectiveness of economic power dispatching.In recent years,with the rapid development of the economy and the improvement of people's living conditions,the proportion of temperature attemperation load in the whole network load has gradually increased,so the grid load is increasingly sensitive to changes in meteorological conditions.How to grasp the law of load change under the influence of meteorological factors and improve the accuracy of short-term load forecasting are the problems that need to be solved urgently.With the advent of big data era,power advanced metering systems and supervisory control and data acquisition systems store massive amounts of load data and meteorological data,explore effective data mining techniques,and correctly analyze changes in load patterns due to meteorological factors,which will become an integral part to imporve the accuracy of short-term load forecasting.Based on the analysis and summary of the existing short-term load forecasting methods,this paper proposes a short-term load forecasting method based on meteorological data and load data,and uses machine learning algorithms to predict and model each load mode.Main tasks as follows:Firstly,the characteristics of power grid load are analyzed.The characteristics of load change are studied from the aspects of daily cycle characteristics,weekly cycle characteristics,seasonal characteristics and meteorological sensitivity.According to its periodic and seasonal meteorological factors,the feasibility of improving the accuracy of short-term load forecasting is explained,and the correlation between temperature and humidity and load at different times in each season was analyzed using the Pearson correlation coefficient method.Secondly,aiming at the problem of inaccurate seasonal load division by fixed month or fixed temperature,an adaptive partitioning method combining k-medoids clustering algorithm and CART decision tree is proposed,and the rules of each season are formulated to obtain corresponding load season mode.By using the non-parametric density fitting algorithm,the seasonal typical daily load curve is extracted from the divided season load curves,and on this basis,the multi-segment load mode is extracted by the important point segmentation to reflect the characteristics of temperature and humidity acting strengths on load differently at different time periods.Then,a new reference day selection algorithm based on pearson correlation coefficient weighting is proposed to filter the reference date,and the screening result is used as the input of the prediction model training set.The short-term load forecasting of each segmented load pattern of the forecasting day is modeld by using the crisscross optimized robust extreme learning machine.After obtaining the prediction results of each segment,it is integrated in time series to obtain the final daily load prediction curve.Finally,the real-time load data,temperature data and relative humidity data of Sydney,NSW,published by Australian AEMO Power Company are simulated.The results show that the proposed load pattern extraction algorithm can be effectively applied to BP neural network,support vector regression and extreme learning machine,which proves its universal applicability.Under the load extraction mode,the robust extreme learning machine based on crisscross optimization has the lowest prediction error,which proves the effectiveness of the prediction algorithm.