Accurate forecasting of annual electricity consumption is of great significance for energy planning and policy making. Given that the literature ignores the impact of feature redundancy and uncertainty of algorithm-specific control parameters of an intelligent optimization algorithm on forecasting accuracy, this paper introduces a max-relevance and min-redundancy(MRMR) algorithm to select the key influencing factors as predictors, proposes an improved Jaya algorithm(iJaya) to optimize the hyper-parameters of support vector regression(SVR) and constructs the annual electricity consumption forecasting model MRMR-iJaya-SVR. Taking the real electricity consumption data of China as an example, this paper validates the forecasting performance of the MRMR-iJaya-SVR. Besides, the yearly electricity consumption data of Beijing are used to test the robustness of the proposed model. The experimental results show that the iJaya algorithm has better global searching ability and is more stable. And the proposed model outperforms benchmark models in both single-step-ahead and multi-step-ahead forecasting. Furthermore, for different datasets, the proposed model has strong robustness.