The Research And Applications Of Time Series Forecasting Methods Based On Machine Learning

Time series forecasting has always been a hot issue which people pay attention to.With accurate forecasting results,people can arrange their work in advance and prevent unfavorable situation,and it is also very important for people to formulate policies.With the continuous progress of science and technology,the time series forecasting method has been developed greatly.At present,the commonly used time series forecasting methods are the traditional time series forecasting method and the machine learning based forecasting method.These methods are easy to use and simple to operate,can get high forecasting accuracy,thus has been widely used in the industry.But when these methods are used in different data sets,the results of the accuracy are large,they are not universal.Therefore,many researchers use combination forecasting methods and hybrid forecasting methods to improve the versatility of these forecasting methods.By combining different traditional time series forecasting methods and machine learning prediction methods together,we can make full use of the advantages of each model,thus improve the accuracy of time series forecasting.In this paper,we propose a new time series forecasting method,BP-SARIMA-ANFIS,which combines back propagation neural network(BP),seasonal differential autoregressive moving average(SARIMA)and adaptive fuzzy neural network(ANFIS).Firstly,the original time series data are predicted by BP,SARIMA and ANFIS respectively,and then the weighted average value of the forecasting results obtained by the three methods is obtained.The weight coefficient plays a very important role in the combination forecasting model.In this paper,the weight coefficient of BP-SARIMA-ANFIS is optimized by differential evolution algorithm(DE).The results of the BPSARIMA-ANFIS are compared with those forecasted results by the other three methods.The comparison results show that the BP-SARIMA-ANFIS method effectively improves the electric load forecasting accuracy of the New South Wales(NSW)in Australia.This paper also proposes a GGNN hybrid forecasting method based on improved gray models and BP neural network.The simulative values of the four improved gray models are taken as the input of BP neural network,and the final forecasting value is obtained by repeated training and fitting.This method uses the genetic algorithm(GA)to optimize the weight and threshold of the GGNN.The effectiveness of GGNN is verified by comparing and analyzing the forecasting results of the GGNN method with the forecasting results of the other six methods.The kernel extreme learning machine(KELM)is a kernel-based learning method,and the kernel function plays an important role in KELM.Different kernel function has different geometric functions,and different chosen kernel functions lead to different generalization ability of KELM.In this paper,we propose a new combination kernel function,which combines the RBF and UKF kernel functions,and apply the new combination kernel function to the KELM(Mixed-KELM).We validate the effectiveness of the Mixed-KELM method by using it to forecast the electric load data of New South Wales,Queensland,and Victoria.Electric load data is affected by many unknown factors,thus often shows a certain degree of randomness.Therefore,before using the Mixed-KELM to forecast the electric load,we first use the empirical mode decomposition(EMD)to denoise the original electric load data.The experimental results show that the EMD-Mixed-KELM has higher forecasting accuracy than RBF-KELM,UKF-KELM and Mixed-KELM.The kernel function plays an extremely important role in the least square support vector machine(LSSVM),which is the key technique of LSSVM.When using LSSVM for classification and regression,choosing an appropriate kernel function is the basis and premise for obtaining a better result.At the same time,this paper also applies the combination kernel function to the LSSVM,and proposes a new EMD-Mixed-LSSVM forecasting method.The EMD-Mixed-LSSVM first uses EMD to denoise the time series,then puts the processed data into LSSVM algorithm which uses the combination kernel function proposed in this paper.The results show that EMD-Mixed-LSSVM can improve the time series forecasting accuracy effectively by comparing the forecasting results of EMD-Mixed-LSSVM method with other methods(RBF-LSSVM,UKF-LSSVM and MixedLSSVM).The main results and contributions of this paper are as follows:(1)A combination forecasting method based on BP neural network,SARIMA and ANFIS is proposed,and the effectiveness of the method is verified.(2)A hybrid forecasting method based on improved gray model and BP neural network is proposed.The accuracy of the algorithm is verified by forecasting the oil consumption in China.(3)A new kernel extreme learning machine forecasting method based on combination kernel function is proposed.The experimental results show that this method can improve the time series forecasting accuracy by forecasting the electric load data of three states in Australia.(4)This paper proposes a new least squares support vector machine(LSSVM)forecasting method based on combination kernel function.The effectiveness of the method is proved by forecasting the price data in Australia.