Research On Reactive Load Prediction Algorithm Based On Attention And Multi-task Learning

In recent years,China’s economy has maintained steady growth.In particular,in the face of the sudden COVID-19 epidemic in 2020,China’s economic growth rate will reach 2.3%,becoming the only major economy in the world to achieve positive growth.With the steady recovery of national social and economic activities,the growth rate of electricity consumption in various industries has gradually begun to be obvious,especially the surge of electricity demand in the information technology industry.In this context,various industries of the society put forward higher requirements on the stable energy supply and intelligent dispatching of the power system.Reactive load prediction is one of the key technologies applied to intelligent power network.Accurate reactive load prediction can provide data support for voltage reactive power optimization,steady-state power flow calculation and analysis,frequency control and other important operations,thus affecting the stability and reliability of the power system.Therefore,how to improve the accuracy of reactive load prediction continuously has become a very important topic in the field of smart grid.Around this topic,this paper combined with the theory and technology of machine learning to conduct in-depth research.The main research content includes three parts:First,this paper analyzes the characteristics of reactive load from the monthly time scale and the daily time scale,and finds that reactive load has obvious trend and periodicity.Based on this analysis and related studies,this paper explores the external factors that affect the reactive load prediction,and finds a certain correlation between the fluctuation of reactive load and date types and meteorological factors.Finally,historical reactive power load,date type,temperature,human comfort index and other characteristics were selected to construct the data set,which is used to train and test the reactive power load prediction model.Secondly,this paper studies the reactive load prediction algorithm based on the Attention mechanism and Long Short-Term Memory(LSTM)neural network.Traditional Attention is generally applied to encoder-decoder structure,and only calculates the weight of the Attention of different time-step vectors,without taking into account the different influences of each feature element inside the vector on the predicted results.To solve this problem,this paper proposes a new calculation method of Attention mechanism,which calculates the Attention coefficients of time-step feature vectors and internal feature elements in input data respectively,and then multiplicates the two results element by element to get the final Attention weight.In this paper,the ALSTM(Attention Based Long Short-Term Memory)reactive load prediction model was constructed by combining the improved ATTENTION mechanism with LSTM,and the experimental results proved that the accuracy of reactive load prediction was effectively improved.Thirdly,this paper studies the reactive load prediction algorithm based on a multi-task learning mechanism.Through qualitative and quantitative analysis of the correlation between reactive power load and active power load in the same time period,it is found that they have a strong positive correlation.However,for the single task reactive load prediction model,the input characteristics cannot include the active load.Mainly to avoid data leakage problems,that is,using future data to make predictions.To solve this problem,a multi-task MT-ALSTM(Multi-task ALSTM)load prediction model is constructed by combining reactive load prediction and active load prediction.Among them,the active load forecasting is the auxiliary task and reactive load forecasting is the main task.The structure of the MT-ALSTM network is mainly based on the multi-task learning method of hard parameter sharing.The data of two tasks and the parameter information of shallow LSTM network are shared to promote the information integration between tasks and achieve the purpose of mutual learning.In addition,a weighted loss function is designed for the MT-ALSTM model,and a two-step grid search algorithm is used to optimize the super parameters.Finally,a comparative experiment is carried out on different time scale test sets.The results show that the MT-ALSTM model not only has a good effect on the reactive load prediction but also has good accuracy for the predicted active load.