Weight allocation of combination prediction based on sequence relative nearness degree

Aiming at the weight allocation problems of combination prediction for a time series, a new method was proposed to evaluate the applicability of the employed models and allocate weights, based on the "nearness" between the test sequence and the corresponding prediction value sequence, which overcame the shortages of existing methods such as mean square error reciprocal weight (1/MSE), entropy weight and optimization weight. The definitions of sequence relative nearness degree (SRND), related sequence trend association and scale interval entropy were given and well discussed, as well as the weight allocation expressions based on SRND. By the example which combined the autoregressive moving average model, functional-coefficient autoregressive model and radial basis function prediction networks in the prediction analysis for the takeoff exhaust gas temperature margin time series, the conclusion is drawn that the prediction accuracy can be effectively improved with the proposed method, compared to 1/MSE and entropy weight methods, while the calculation mount is far lower than optimization weight method.