Very Short Term Load Forecasting Based On Echo State Networks

Electricity load forecasting is a basic task for power grid management, which can make a direct effect on the security and economic operation of power system. Very short term load forecasting (VSTLF), dealing with a forecast horizon of one to several dozen minutes ahead, is of special importance in real time economic dispatch, Automatic Generation Control (AGC) and the load frequency control. Accurate forecasting will be helpful to the safety of power system, the balance of supply and demand and improve the eligibility rate of power system frequency. Therefore, the study of VSTLF will be of important practical significance.The main content of VSTLF and load forecasting are introduced. First, this paper briefly introduces the classification of load forecasting, load characteristics and load model. Thereafter the basic content of VSTLF was expounded, which includes load data preprocess, the selection of forecast samples and the usual prediction method. Finally, the reasons for forecasting error and some evaluating indices for forecasting are also introduced.A very short term load forecasting method based on robust regression and echo state networks is proposed considering the very short forecast time of VSTLF. As a recurrent neural network, echo state networks has a dynamic reservoir as the hidden layer and trained by linear regression, which can model any given function and makes it training very quickly. As a result, echo state networks can meet the needs of VSTLF. In addition, the using of robust regression to training can effectively reduce the effects of outliers and promote the accuracy of forecasting.According to the outline of VSTLF, the simulation analysis is implemented. Considering the different condition of exception data handling during training and forecasting, different data preprocess methods are used for the two stages. In order to optimize the selection of input, the forecasting sample is selected based on a shape similarity criterion and three kinds of network input are builted. Three kinds of input are compared through the correlative analysis and simulation. Combing with data preprocessing and network input selection, three forecasting methods based on BP network, echo state networks and robust echo state networks are implemented. The results of calculation example show that the method based on robust echo state networks has the advantage of shorter training period and higher forecast accuracy, being a feasible and effective forecasting method.