Data-driven Reconstruction Of Missing Data And Power Load Forecasting

Power load forecasting is the basis for formulating power supply plans and ensuring the balance between power supply and demand.Improving the accuracy of power load forecasting will help improve the operation quality and economic benefits of the power system.To obtain high-precision load forecasting results,sufficient and high-quality historical data is a necessary prerequisite.However,due to SACDA equipment failures,signal transmission disturbances,weather influences and personnel misoperations,etc.,the partial loss of power load data or the integration of erroneous data will affect the quality of the data.The integration of external factors such as weather provides a direction for the accuracy of load forecasting.The development of deep learning technology provides technical possibilities and technical means for the introduction of such external information into neural networks to implement effective intelligent forecasting.This paper discusses the ways to cope with and solve the problem of a variety of power load data missing(especially when the missing rate is large),and uses a variety of deep learning algorithms to carry out the research work of system load forecasting.Because the minimum time scale of the load data used is 1h,only the forecasting effect of the established forecasting model is verified for the next hour and the next day.This article first conducts the research on the missing processing of original historical data.Different degrees of missing rates are set for historical power load data,and several methods based on statistical values,machine learning and deep learning are used to fill and reconstruct the data,and the effects of different methods are compared through experiments.At the same time,the reconstructed data is used to predict the future load to further verify the accuracy of filling the missing data and the quality of the filled data.Then,conducted a variety of deep neural network application research.The forecast quality of hourly load forecasting of three forecasting models based on multilayer perceptron neural network,cyclic neural network and long-and short-term memory neural network is compared and studied.The empirical mode decomposition method is also used to decompose the original load sequence into multiple sub-sequences,which are combined with the long and short-term memory network and random forest respectively to explore ways to speed up model learning and obtain higher prediction accuracy,and conduct experimental verification.Finally,external information such as temperature,humidity,and wind speed is introduced,and multi-source data constructed using date,weather,and historical load data are used to select the best input feature using the Xgboost algorithm.On this basis,the accuracy of single model and hybrid model in short-term load forecasting is compared.