Short-term Power Load Forecasting Based On Improved Generalized Regression Neural Network

Power system load forecasting is the basic content of daily work for power dispatching department at all levels of State Grid Corporation of China,it has become an important index to evaluate the modern operation and management of power system,especially accurate short-term load forecasting is more instructive.Under the condition that power grid is becoming more and more perfect,traditional prediction method is very difficult to improve prediction accuracy.This paper gives a brief overview of the power load forecasting,introduces the characteristics and classification of load forecasting,discusses the steps of the prediction and the error evaluation index,and describes the Generalized Regression Neural Network(GRNN).The main contents and work are as follows:(I)Taking Baiyin area as an example,the characteristics of short-term load in this area are emphatically analyzed.In the meantime,collecting and sorting the load data and meteorological information in this area,studying the change law of load and discussing the relationship between load and its influencing factors.As the content of load data and influencing factors are more,the correlation analysis method is considered to simplify the influence factors,extract the information related to the load and remove the unrelated factors,thus reducing the workload for the establishment of the prediction model.(II)According to the determination of load characteristics and related influencing factors,a prediction model of neural network is established.For this reason,we use Error Back Propagation Neural Network(BPNN)and GRNN to predict.Compared with BPNN,the function approximation ability of GRNN is very strong,and the factors that need to be adjusted are less,the training only depends on the input sample,and it can prevent the adverse effects on the result.After forecasting the integer point time of three days,we find that the prediction error of GRNN is not only smaller than that of the existing prediction system,but also more accurate than that of BPNN.(III)Swarm Intelligence Optimization Algorithm is a new optimization method in recent years.This paper introduce the principle and steps of the Fruit fly Optimization Algorithm(FOA),10 test functions are selected to study the influence of each parameter on this algorithm,and a general conclusion is drawn.Although the performance of FOA is excellent,it is easy to fall into the local optimal,and the stability is not high.Therefore,starting with one of the parameters of FOA-search distance l,this paper puts forward two improved methods.The first is to reduce the l gradually and the second is to make the l increase first and then decrease.The simulation results of the test function show that the second improved methods not only have better convergence accuracy than the original ones,but also are better than the original ones in the stability of the algorithm.(IV)The optimization training of GRNN is actually to optimize its smoothing factor ?.In this paper,we use Particle Swarm Optimization(PSO),FOA and IFOA(Improved FOA,the second improved methods)to optimize the ? of GRNN and predict the integer point time load of three days.It is proved that the prediction error will be further reduced by using PSO or FOA to optimize the GRNN,but Improved FOA to optimize GRNN is the least error in the method used.