Abstract
Short-term load forecasting (STLF) enables distribution system operators to perform efficient energy management by flexibly engaging energy consumers under the intelligent demand-response program in the smart grid (SG). This paper develops a fast and accurate hybrid load forecasting model with this motivation. The proposed model integrates a locally weighted support vector regression (LWSVR) based forecaster with two modules. These modules are feature engineering (FE) and adaptive grasshopper optimization (AGO) based optimizers. In the FE module, firstly, the proposed hybrid feature selector (HFS) is developed using wrapper and filter techniques to determine an optimal subset of features. Furthermore, the instance-based Relief-F (REF) and information theoretic-based mutual information (MI) filters are used to decrease the curse of feature dimensionality by finding and eradicating extraneous features. In addition, to overcome the overfitting problem, the HFS module is further optimized using the recursive feature elimination (RFE) wrapper feature selection technique. The essential features are then extracted using a radial basis Kernel-based principal component analysis (RBF-KPCA) algorithm to eliminate the dimensionality reduction problem. The AGO algorithm tunes the LWSVR model’s appropriate parameters to effectively evade entrapping into local optimum and yield accurate prediction results. However, the efficacy and productiveness of the forecasting model are differentiated equally by its convergence rate and stability. Actual hourly load data of two states of Australia (New South Wales (NSW) and Victoria (VIC)) and California Independent System Operator United States (CAISO-US) are employed as a case study to estimate the effectiveness and applicability of the designed model. Empirical results show that the devised model surpasses benchmark models (single and hybrid) in terms of stability, accuracy, and convergence rate.
•A Hybrid Framework for Short Term Load Forecasting is presented.•Hybrid feature engineering, optimization and forecasting modules.•Instance-based Relief-F and mutual information feature selection.•Radial basis Kernel-based principal component analysis to eliminate dimensionality problem.•Adaptive grasshopper optimization is used to tune support vector regression.