Research And Application On Combined Forecasting Models

This paper has used multiple linear regression model and GM(l,n) to forecast the oil consumption. Taking into account GM(l,n) is not suitable for complex nonlinear function approximation. So we use a gray neural network (GNNM(l,n)) model, which integrates the artificial neural network, thus will avoid the single model’s weaknesses, thereby improve the model system’s performance. However, from the point of gray neural network training process, based on the shortcoming of "no further amendments to the parameters" about GNNM(l,n), we have to apply genetic algorithm to optimize the parameters, that is GA-GNNM(l,n). Finally, several models are applied to the oil consumption forecasts so as to achieve the effectiveness of the model test.However, a single model does not fully reflect the information and variation of oil consumption, therefore combined forecasting model has been becoming a new research trend. This paper introduces the minimum error sum of square as well as the variable weights of combination forecasting model, and in view of the weight value ranges, we use a traditional combination theory and a traditional combination method based on the proposed no negative constraint theory. In order to improve the prediction accuracy, we combine every forecast for a total of 26 combination models using the traditional combination method, which will avoid the individual risk prediction methods appearing "over-fitting" to reduce the prediction accuracy. The combination forecasting errors about MAE、MSE、MAPE、SSE are also compared by the traditional combination method and the no negative constraint method among the 26 combinations models of the oil consumption. We can also draw a conclusion from the weight coefficient table and the error tables that whether the single method should be eliminated, which will improve the reliability of forecasting. What’s more, compared with the traditional forecasting combination, the forecasting accuracies improve when negative weights are assigned. Therefore, we believe that the combination method of TCM-NNCT is feasible and effective, without the weight of the value to [0,1],only need to meet ∑li=1.