Ultra-short-term Power Load Forecasting Model Based on MIC and MA-LSTNet
In multivariate time-series ultra-short-term power load forecasting,long-term and short-term time patterns of-ten exist among variables.The long-and short-term time-series network(LSTNet)can extract short-term fluctuations and long-term trends between weather factors and loads,which improves the forecasting accuracy.In this paper,we es-tablished an ultra-short-term power load forecasting model based on maximum information coefficient(MIC)and long-and short-term time series network with multi-head attention mechanism(MA-LSTNET).Firstly,we utilized the maxi-mum information coefficient(MIC)to analyze the correlation between weather variables and forecast series in each load lag period.Then,we employed the symbol aggregation approximation(SAX)to quantify the correlation curves and perform optimal selection of weather variables to minimize model input redundancy.Secondly,we improved the LST-Net,and proposed a long-and short-term time series network using multi-head attention mechanism(MA-LSTNet).By integrating a self-attention layer into the nonlinear part,the model is able to extract non-seasonal and non-periodic long-and short-term time patterns.So far,compared with other models,the model proposed in this paper exhibits the best predictive performance.
long-and short-term patternmaximum information coefficientrecurrent-skip layerattention mecha-nismsymbol aggregation approximation
万方数据
维普
