Robust electric load forecasting through ensemble learning: A stacking approach with empirical mode decomposition and transfer learning

Abstract

Recent advancements in artificial intelligence (AI) have significantly influenced various disciplines, including electricity demand forecasting within power systems. This study introduces a methodology that emphasizes on predicting total energy consumption rather than limiting the scope to specific sectors. By integrating Empirical Mode Decomposition (EMD) with Transfer Learning (TL), the proposed model enhances the accuracy and generalization capability of ensemble models. The methodology achieves this by decomposing input data features into linear and nonlinear components, that optimized the resource allocation, encourages to use simpler models, and mitigating overfitting risks. TL further strengthens the model's adaptability, allowing it to accommodate diverse load patterns from multiple sectors. This adaptability facilitates the integration of sector-specific load into a comprehensive framework, leading to more accurate predictions of net load demand for power station operations. Experimental evaluations validated the model’s superior performance, achieving a Mean Absolute Error (MAE) of 82.58, a Root Mean Square Error (RMSE) of 95.375, and a Mean Absolute Percentage Error (MAPE) of 1.06 %, this contributes 2.85 % improvement over conventional methods. The proposed model further validated on the widely utilized New South Wales (NSW) dataset revealed an MAE of 96.23, an RMSE of 102.16, and a MAPE of 1.02 %. A predictive accuracy of 98.98 % was achieved using the proposed model, which outperforms state-of-the-art models like N-BEATS and DLinear and other advanced ensemble techniques. Statistical tests, such as the Friedman test and Nemenyi post-hoc analysis, confirm the strength of the proposed model, regularly placing it as the top performer among the other methods. This enables the proposed model to be applicable in the real world by predicting the energy consumption at a broader level rather than at a sector level. Moreover, the proposed model’s outcomes illustrate, that the framework is reliable and generalizable in nature which leads to better resource optimization and promotion of energy efficiency practices in load forecasting can be achieved.