An integrative approach to enhance load forecasting accuracy in power systems based on multivariate feature selection and selective stacking ensemble modeling

Accurate load forecasting is important for the safe and stable operation of power systems. Although there are many load forecasting methods, most of them focus on using a single model to forecast a single load time series, ignoring the limitations of a single model and the influence of other factors on the load. To address this problem, this paper proposes a load forecasting method that combines typical correlation analysis (CCA), stacking ensemble forecasting, and intelligent algorithm optimization. First, the CCA algorithm is utilized to screen the influential factors with high correlation from the multivariate meteorological factors; second, a 5-fold cross-validation algorithm is used to reconstruct the data, and based on the prediction results, 5 models are selected as the base learners from 10 mainstream prediction models. Then, the Generalized Regularized Extreme Learning Machine (GRELM) is used as the meta-learner and its parameters are optimized using the Supply and Demand Optimization (SDO) algorithm. Finally, the model was used to fit and summarize load data from Panama over a two-year period through four sets of experiments and multiple evaluation metrics. The results show that the proposed model has an average RMSE of 20.63, MAE of 15.15, R of 0.9937, and MAPE of 0.0124 in the four experimental datasets, and the average RMSE of the proposed model in this study is higher than the control model by 10.41 % to 29.55 %, and the average MAE is higher by 11.79 % to 29.14 %. These results indicate that the CCA-SDO-Ensemble prediction model proposed in this paper has higher prediction accuracy.