Two Types Of Combination Forecasting Methods And Their Applications

As a widely discussed issue, there are already a large number of models to solve time series prediction problems. These models can be roughly divided into two cate-gories:linear and nonlinear ones. Linear models such as ARMA, ARIMA, SARIMA models can adjust itself according to the new obtained data so as to improve the fore-casting accuracy, however, they can only distinguish the overall trend of the data but not all of the factors which effect the data changes; ARCH and GARCH models can not only get the conditional variance, but also the iteration method of the predictive value, and shall calculate the risk predictive value according to the iterative method, but in order to ensure non-negative, it is usually assumed that all coefficients in these models are greater than0, any lagging terms implied by these constraints will increase and thus it rules out the random fluctuations in behavior, which may come up a shock phenomenon when models are estimated. Nonlinear models such as neural networks, have the advantages of fast computing power, and the ability to approximate arbitrary continuous mapping, powerful learning ability and dynamic analysis capabilities, they have been widely used in the regression, however, they are easy to fall into local min-imum and over-fitting; support vector machine algorithm converts practical problems to high-dimensional feature space through a nonlinear transformation, and constructs linear discrimination function in high-dimensional space, this algorithm not only pro-vides a fine generalization ability but also solves the high-dimensional problem, which makes the algorithm complexity is independent with the dimension of the sample, but this model is very sensitive to the selection of parameters.Taking into account the advantages and disadvantages of these various prediction models, Bate and Granger proposed the idea of combination forecasting in1969. This idea and its extension bring more and more choices to overcome the above-mentioned defects of an individual forecasting model. This paper is tightly around a combination idea based on the combination of forecasting methods for research, which is so dif-ferent from the widely combination idea based on combination of forecasting results. This idea gives full play to the strengths of the various methods. Based on this idea, this paper develops two types of combination methods:1) the original time series trend and random fluctuations are predicted by different proper models respectively, and then combine their predicted values to form the final forecasts;2) the original time series is first decomposed into some different frequency sub-series by nonlinear signal process-ing technology and then the resulting high frequency and low frequency series, then these two series are predicted by the same or different models, and the combination of their forecasts is just the final prediction results.On the other hand, electricity load and price forecasting have very great practical significance to maximize the social benefits, but their accuracy forecasting are always difficult tasks. Thus this paper uses these two tasks as the application background for the above-mentioned new combination forecasting. According to the above-mentioned methods, this paper establishes three forecasting models:1) Under the first approach, taking into account the multiple seasonality of electric load data, the seasonal ARIMA model (SARIMA) first models and predicts the original time series, which can obtain the residual series (i.e. the random fluctuations), then due to the highly nonlinear na-ture of this residual series, the nonlinear model BP neural network model is introduced to predict it, and finally the predicted values of the original series and residuals are combined to get the final prediction results;2) Also under the first approach, taking into account the strong volatility of electricity price data, the original time series is first predicted by the generalized autoregressive conditional heteroscedasticity model (GARCH) which is the mainstream algorithm for processing the series contained vari-ance gathered effect, then its residual series is modeled by support vector machine (SVM) which has outstanding generalization ability, and the ultimate combination pre-dicted results can be obtained in the same way of the above model;3) Under the second approach, the original series is decomposed into some different frequency sub-series by empirical mode decomposition (EMD), then the higher and lower frequency sub-series are summed up as the high frequency and low frequency series respectively, due to these two series contained far less fluctuation factors than the original series, they can be simulated by the forecasting models more conveniently, and this paper selects the BP neural network to forecast these sub-series and then obtains the combined results. In addition, this paper establish an effective degree for each models which can measure their effectiveness, and the experiments of power load and electricity price forecast of Australia prove these methods are effective to improve the prediction accuracy.The main research achievements and contributions are as follows:1) This paper develops the idea of the combination forecasting through the intro-duction of two new combination forecasting methods and the resulting three prediction models which is derived from the above two methods, in addition, this paper establish an effective degree for each models which can measure their effectiveness; 2) Electricity load and price forecasting are two hot and also difficult issues of the electricity market. This paper attempts to find the highly accurate forecasting models by theoretical analysis and experiments for them, and improve the forecasting accuracy ultimately, which is a very meaningful attempt.