Research On Short Term Load Forecasting Based On Recurrent Neural Networks

Power load forecasting is one of the key parts in the daily operation of the power system.Precise and effective forecasting enables system operators to make reliable unit scheduling and power allocation decisions,which is helpful to reduce the consumption of non-renewable energy,lower the cost of power generation,and promote the sustainable development of society.As a classic forecasting problem,power load forecasting has received great attention from researchers.In recent years,the continuous development of artificial intelligence technology has provided a series of effective methods for power load forecasting.Considering the temporal feature of power load,this paper proposed a point forecast method using long short term memory neural networks,which is based on empirical mode decomposition and particle swarm optimization.To be specific,empirical mode decomposition was used to decompose the original power load series into a combination of several intrinsic mode components and components with relatively stable and consistent frequency.Then,for each component,the long short-term memory neural networks was constructed to capture the time series feature of power load.Considering that forecast performance can not be guaranteed when parameters of long short-term memory neural networks is initialized randomly,particle swarm optimization was applied to search the optimal initial parameters.Case studies based on real dataset show that the above forecast model of empirical mode decomposition and particle swarm optimization can get an get a Mean Absolute Percentage Error of 1.43% when forecast day-ahead load power.Besides,the above forecast model can improve load power forecast accuracy of more than 1% compared with recurrent neural networks and gated recurrent unit neural networks.Meanwhile,the forecast error on the whole test set is less than 3%.To capture the uncertainties of future load,the interval forecast method was developed based on long short-term memory neural networks.First,the data cleaning method similar to the above point forecast model was used to preprocess the historical data of load power.Then,an interval forecast model based on long short-term neural networks was constructed to estimate the lower and upper bounds of load power.Finally,the recently proposed the multiple threshold sets based multi-objective particle swarm optimization algorithm sets was developed to tune the parameters of the interval forecast model.Case studies show that the multiple threshold sets based multi-objective particle swarm optimization algorithm can effectively improve the forecast performance of long short-term memory neural networks,in which the hypervolume of final Pareto-optimal fronts is 0.5408,6.67% and 8.59% larger than time-varying based multi-objective particle swarm optimization algorithm and dominance time-based the multiple threshold sets based multi-objective particle swarm optimization algorithm.