THE GAs-BASED APPROACH TO AGGREGATION-DISAGGREGATION OF ARMA AND ITS APPLICATION IN THE LOAD FORECASTING OF ELECTRICAL POWER SYSTEM

The simulation and optimizing control for large-scale complex systems are depended on the models which can describe the system hierarchical organization structure and dynamic behavior appropriately and accurately. In view of some deficiency of traditional modeling methods, a new approach to system modeling? multiresolution modeling and aggregationldisaggregation?has been developed since 1980, and is now being used in the simulation and optimizing decision-making for the large scale complex systems. In this paper, the characteristics of the large-scale complex systems and their hierarchical embodying relationships are discussed at first. Then, the advantages of Multi-Resolution Modeling (MRM) over the traditional modeling methods are reviewed in detail in the first chapter, and the principle of time series modeling is discussed in the 2nd chapter. Thereupon, in the 3rd chapter, some important concepts such as variable granularity analysis, aggregation and disaggregation, smooth and consistency transformation between different resolution models are studied so as to propose certain implementing approaches. As a result, an example of simulation implementation for the multiresolution modeling and forecasting of time series is given to illustrate the efficiency of MRM. In the forth chapter, the principle of multiresolution ARMA modeling is researched in depth. And then, a kind of method for determining the order and integrating the parameters of aggregated ARMA model from higher resolution ARMA model is suggested. The simulating result by the data of a practical system shows its efficiency. In the fifth chapter, the problem of model disaggregation is stated at first. Then, a new way to disaggregate coarse granularity data and lower resolution model is discussed, through which the order and AR parameters of the disaggregated ARMA can be induced firstly to optimize the disaggregating coefficient matrix of coarse granularity time series. Thereby, the variable resolution disaggregation of ARMA models can be carried out through the optimizing index of order and AR parameters of disaggregated ARMA, and an algorithm of ARMA disaggregation based on GAs is proposed, which is testified to be effective through the simulation. In sixth chapter, as an example of applying multiresolution modeling and aggregationldisaggregation methods to practical systems, a new approach is suggested for the load forecasting of electrical power system(EPS). From the viewpoint of signal analysis, the load curve of EPS can be considered as a linear combination of lower frequencies part and higher frequencies part. Thus the total load curve can be decomposed as the low frequencies term of historical trend and the high frequencies term of stochastic time series by Butterworth filter. Then, the Radial Basis Function neural network is employed as a tool to be trained by the data of historical load trend and the corresponding temperature records so as to predict the trend part of the future load, while the multiresolution models of ARMA are built with the data of stochastic time series of EPS through aggregationldisaggregation methods, which can be used to predict stochastic part of the future load. Finally, the total load of EPS can be forecasted by summing the two parts of the predicted values of both the static trend and stochastic distributed terms. The approach has been tested to be satisfied by the load data of a practical EPS. The results show that the application of aggregationldisaggregation in load forecasting is successful.