一种基于动态关系辨识算法的短期预测方法

A Short-term Forecasting Method Based on Dynamic Identification Algorithm and Its Application

针对目前众多短期预测问题中存在的业务数据波动性增大、预测结果影响因素增多等问题,结合自适应建模技术,提出了一种基于动态关系辨识算法的统计类短期预测方法。该方法首先根据预测问题提出动态关系模型,然后结合不断更新的业务数据和最佳预测精度值,对模型参数和模型结构系数进行动态调整,从而使得决策者能够始终基于模型的最优状态获得预测结果。研究结果表明:针对波动性较大的业务数据,该方法自我调整能力较强,对近期数据的跟踪性较好,鲁棒性强,易于推广应用。

In short-term forecasts, the predicted values are usually treated with monthly, weekly or even daily periodicities. The short- term forecast has played more and more important roles in people＇s life and work because of rapid development in information technology. Short-term forecasts have many features, such as shorter forecasting period, more elusive regularity of practical data, many factors influencing the forecasting results, and so on. These features increase the difficulty of prediction. In order to reduce decision risks, decision-makers need to seek proper forecasting tools to reveal the regularity of the predicted data and make a reliable and accurate forecast. A variety of models and algorithms on short-term forecasts are proposed in recent years. Among them the traditional statistical technologies are still the most popular methods because they are easy to be accepted and applied by decision-makers. Statistical methods generally contain two kinds of models ： static models and dynamic models. The static model has a fixed structure which usually reduces its ability to trace constant changes in the environment. The dynamic model is constructed based on the analysis of certain stochastic process. Compared with the static model, the dynamic model can produce more accurate and credible results and has the potential defect of low robustness in practical problems. After including the features of current short-term forecasts as well as the imperfection of existing statistical forecasting methods, a new short-term forecasting method based on the dynamic relation identification approach is presented in the paper. Its forecasting process is briefly described as follows. Firstly, the dynamic relation model reflects the relational pattern between the forecasting value and its correlative influencing factors. Secondly, in each forecasting period the optimal forecast precision is calculated by the precision- determined formula and the newly actual data. Thirdly, both the parameters and structural coefficients of the dynamic relation model can be adjusted based on the optimal forecasting precision. Finally, the optimal forecasting results are produced by the updated dynamic model which extracts the trend of the predicted data more reasonably and accurately. The proposed method is tested by a practical case in which high accurate forecasting results are obtained. Compared with other statistical forecasting approaches such as the moving average and the exponential smoothing, our method＇s accuracy is higher by about 30%. In the case, the proposed method also demonstrates good robustness and adaptability to the constantly changing circumstance. The method has other advantages inshort-term forecasts. For example, the capability of tracing different-period data is guaranteed by the forecasting factor. The recursive formulas in the method are so normalized that they can be programmed easily and the multiple- period-ahead predictions can be provided. These characteristics of this method improve the ability of dealing with actual forecasting problems and facilitate effective decision-making processes.