Long-Term Workload Forecasting in Grid Cloud using Deep Ensemble Model

Smart grid is the direction and trend of power industry development. Cloud computing, as a basic component of smart grid information platform, provides flexible virtual resources for grid system information application. However, unreasonable allocation of computing resources will lead to resource waste and high resource load. Resource prediction can solve this problem well by dynamically allocate system resources according to the predicted value of resources. This paper presents a simple and efficient ensemble model resource prediction method in gird cloud environment. We first train three kinds of novel deep learning time series prediction models, namely N-BEATS, N-HITS and TFT to build ensemble model. Then a simple error based dynamic model selection method is used to choose best model in prediction. Compare with the single model prediction result in ShanDong Grid dataset, the accuracy of proposed ensemble model is higher than best single model about 37%, 34%, 36% in CPU, memory and disk load forecasting respectively. Meanwhile, the proposed method has strong generalization ability, and it is not necessary to train a separate prediction model for each information system. The mean absolute percentage error of ensemble model in CPU, memory and disk data is only 9.9%, 0.68% and 0.14% respectively.