Research On Online Prediction Of Time Series Based On Improved Kernel Least-mean-square Algorithm

Time series are discrete data arranged in chronological order of a certain statistical value,which is widely used in the fields of environment,engineering,production and life.Time series prediction can discover and refine the connections and laws within or between things,and provide decision-making basis for human production and life,which has important practical significance.With the rapid development of current digital technology,time series show the following characteristics: large amount of data,high dimensionality,time-varying and noisy.This puts forward higher requirements for online time series prediction.This paper is based on the kernel least mean square method to study the online prediction of time series.Aiming at the time series containing noise or outliers,non-stationary and time-varying,the kernel least mean square method is improved to reduce the computational complexity and improve the adaptive adjustment ability.Enhance the ability to track time-varying characteristics,thereby improving the prediction accuracy.The kernel least-mean-square method is a classic method in the field of time series online prediction.First of all,as the data volume increases linearly,model calculation burden and memory requirements continue to increase,and time series often have unknown noise or outliers,a quantized kernel least mean square method based on the sequential outlier criterion(QKLMS-SOC)is proposed.The method uses the sequential outliers criterion to eliminate the outliers of the sample data,and uses the online vector quantization method to distinguish the redundant data.While discarding the redundant data,the information is extracted to update the network coefficients,thereby forming a compact and accurate model.Secondly,to improve the model from the perspective of sparseness,adaptive sparse quantization kernel least-mean-square(ASQ-KLMS)algorithm is proposed.Use the sequential outlier criterion to ignore outliers to prevent them from being accepted as new centers,and add the coherence criterion to the quantized kernel least mean square algorithm to reduce the size of the dictionary,thereby reducing computational complexity and increasing weight adaptation to improve the noise immunity of the model in a time-varying environment.In addition,in view of the problem that the kernel least-mean-square method cannot capture the time characteristics of nonlinear dynamic systems,combining it with the neural network method,quantized kernel least-mean-square based on echo state network(ES-QKLMS)algorithm is proposed.This method integrates the reserve pool of the echo state network with the kernel least-mean-square,which improves the ability of the model to track time characteristics.The vector quantization method can suppress the growth of the model dictionary,thereby improving the calculation efficiency of the model.In this paper,online prediction simulation experiments are performed using the Lorenz dataset with Gaussian white noise,the El Ni?o and Southern Oscillation dataset,and the Beijing Air Quality Index dataset to prove the effectiveness of the method proposed in this paper.