Grid can select, share and aggregate geographically distributed and heterogeneous resources as an ensemble to solve large-scale scientific, engineering and commercial problems. In grid environment, management of grid load is a complicated and important problem. Selecting the proper resources from the resources pool which varies dynamically is an important task for grid scheduler. In order to select grid resources, scheduler must obtain grid resources information which is mainly from grid resources monitoring and gird resources prediction. Traditional scheduler makes schedule strategies through the current information obtained by grid resources monitoring. Combining the current grid resources information and future information from prediction, the grid scheduler will make better schedule strategies. Gird resources prediction will offer important information for scheduler and it can make the usage of grid resources more reasonable and improve the efficiency of the whole system.Currently, the models of grid resources prediction are mainly traditional statistical models(MA,AR,ARMA,ARIMA) which works well in linear time series prediction. However, grid resources time series are often nonlinear. In this case, the prediction accuracy of these models will decline. With the development of Artificial Neural Network (ANN), ANNs are applied to model time series prediction. The parallel, nonlinearity and robustness of ANNs make ANNs work well in nonlinear time series prediction. And Back Propagation Neural Network(BPNN) is the most popular ANN. ANNs have several shortcomings:(1) The structures of ANNs are hard to determine;(2)The number of training samples required by ANNs is too large;(3) ANNs are easy to over learn.(4) ANNs are easy to trap into local minima.Support Vector Machines (SVMs) overcome effectively the shortcomings of ANNs. SVMs are used extensively to model time series prediction due to their superb generality ability, the only solution on convex set and the sparseness of solutions. This study selected Support Vector Regression (SVR), which is one version of Support Vector Machines, to model grid resources prediction. This study introduced detailedly the development of SVMs, the basic thinking of SVMs, the geometry explanation of SVMs, the advantages of SVMs for modeling nonlinear time series prediction, the reasoning process of linear and nonlinear SVMs and the theory of kernel theory.Support Vector Machines (SVMs) overcome effectively the shortcomings of ANNs. SVMs are used extensively to model time series prediction due to their superb generality ability, the only solution on convex set and the sparseness of solutions. This study selected Support Vector Regression (SVR), which is one version of Support Vector Machines, to model grid resources prediction. This study introduced detailedly the development of SVMs, the basic thinking of SVMs, the geometry explanation of SVMs, the advantages of SVMs for modeling nonlinear time series prediction, the reasoning process of linear and nonlinear SVMs and the theory of kernel theory.The SVRs'free parameters influence greatly the prediction performance of SVRs. There is, nevertheless, no general method for determining SVRs'parameters, which is an important reason for limiting the development of SVRs. Trial and error method is the most popular method to determine SVRs'parameters. But this method is time-consuming, far from efficient, not stable and need some luck on many occasions; Therefore, this method cannot satisfied the requirement of grid resources prediction.Genetic Algorithm (GA) is based on natural evolution theory and genetic theory and it is one kind of bionics algorithm which solves complicated optimization problems. GA is applied extensively to combinatorial optimization, self-adaptive control, machine learning and programming design owing to its versatility, stability and globally searching ability. This study proposed the GA-SVR model which used GA to optimize SVR's parameters, then used the optimized model to predict grid resources prediction so as to improve the efficiency and accuracy of grid resources prediction.This study introduced the principle and characters of GA detailedly. There are a variety of GAs, among which we should choose the one which is most suitable for optimizing SVR. This study described three popular kinds of GAs: (1) traditional binary-coded GA (BGA); (2) Real-coded GA with Heuristic Crossover and Uniform Mutation (HRGA); (3) Real-coded GA with Simulated Binary Crossover and Polynomial Mutation (SRGA); During the process of GA optimizing SVR, it is found that SRGA cannot cope with the bounded searching space well and the children generated are possible to be out the searching bound. However, Optimizing SVR is just such a problem. This study, hence, proposed an improved SRGA (ISRGA) based on statistical theory. The ISRGA can deal with the bounded searching space well in order to make every child generated within the searching bound. It proves every child within the searching space to improve the efficiency of optimizing SVR. The ultimate target of this algorithmic improvement is to improve the efficiency of grid resources prediction. The SVR prediction models optimized by four kinds of GAs are named BGA-SVR, HRGA-SVR, SRGA-SVR and ISRGA-SVR. This study illustrated the architecture and principle of GA-SVR. At the same time, cross validation technique is applied in order to decline the dependency of parameters on training samples.This study simulated grid resources prediction process. In order to make the simulation results more versatile and comparable, the experiments used the international benchmark data set. Our experiments compared the performance of grid resources prediction of BPNN, SVR whose parameters are optimized by trial-and-error procedure(T-SVR) and GA-SVR (including BGA-SVR, HRGA-SVR, SRGA-SVR and ISRGA-SVR). The experimental results showed that GA-SVR worked better than the other two and BPPNN worked better than T-SVR. It means that GA-SVR worked better than BPNN and trial-and-error method indeed needs some luck. In additional, the experimental results showed ISRGA-SVR worked best among the four kinds of GA-SVRs. It means that ISRGA-SVR is most suitable for grid resources prediction among the models in our study. Also, it means the improved SRGA is successful in optimizing SVR.Our study will boost the development of grid resources prediction to some extent. The future work is as follows:(1) We will study least squared SVM (LSSVM) and attempt to apply it to grid resources prediction.(2) We will study some other optimization algorithms such as Particle Swarm Optimization (PSO) to optimize the prediction models.(3) We will study some parallel optimization algorithms such as parallel GA and parallel PSO, which can be more suitable for the distributed structure of grid, aiming to improve the efficiency of parameters optimization. This direction will focus on dividing the searching space so as to parallel optimization algorithms... |