Bayesian Model Averaging Base On The Predictive Likelihood And Using It In Power Load Forecasting

Since 1969 when Bates and Granger proposed combination forecasting,it has made great progress.The combination forecasting combines various individual methods to take full advantage of useful information. So that it can improve the prediction accuracy. The key to the combination forecasting is to determine the weight of each individual forecasting model and it can be divided into fixed weight combination forecasting and changeable weight combination forecasting, judging from its time-defined changeability. At present,We have do a lot research on the fixed weight combination forecasting, and we have already has many mature methods on it. However, the fixed weight combination forecasting is difficult to reflect the changeable behavior of individual models. On the contrary, the changeable weight combination forecasting can do it so as to improve the model prediction accuracy and enhance the practicality of the forecasting model. But it is hard to find a way to determine its weight because its weight is a function of the time. The Bayesian method can clearly show the information update process, it can also combine the subjective information and data with various kinds of interventions. Therefore, the Bayesian combination forecasting methods has great significance in adaptability to update the combining weights.There are two main ideas on Bayesian combination forecasting method:the first idea is adopting the combination forecasting form of the weighted average, and then explaining the weights with Bayesian probability rule. The second idea is discussing the how the prior distribution of the forecasted variable was update by the individual forecasting models. i.e. obtaining the optimal form of combination and optimal weights.Bayesian Model Aversging can reduce the model and parameter uncertainty.The key point to carry on combination forecast with BMA is the posterior probability.We have extended the standard approach to Bayesian forecast combination by forming the weights for the forecast combinations form the predictive likelihood rather than the standard maginal likelihood.The new approach overcome the marginal likelihood is over-reliance on the prior information and have small sample properties by split the sample to two parts.Finally,we using the new approach to predict the power load forecast.The result show that the model that performance well would get more weight,the weight and the model selection is asymptotic.The new approach performance well than others.