An Autoencoder-Based Support Vector Machine Optimized with Bayesian Optimization for Net Load Forecasting with Uncertainty Estimation

Load forecasting stands as a crucial factor in the economic and reliable operation of power systems. With the increasing integration of Renewable Energy Sources (RES) into power systems, Net Load Forecasting (NLF) emerges as a significant concern in power systems operations. However, the adoption of RES introduces challenges, notably uncertainties stemming from environmental conditions, thereby rendering NLF a critical issue. This study presents an Autoencoder-based Support Vector Machine optimized with a Bayesian optimization (ASBO) approach for NLF in a time series framework, leveraging an Autoencoder for feature extraction from input data. Specifically, Bayesian optimization is employed to fine-tune SVM hyperparameters, directly influencing forecasting performance. Finally, the proposed method is employed to estimate the uncertainties posed by RES. Various evaluation metrics are utilized to analyze prediction accuracy and uncertainty estimation. To validate the efficacy of the proposed method, historical time series data from Austria spanning 2018 and 2019 are utilized, encompassing parameters such as temperature, actual loads, solar energy, and wind turbine generation at hourly resolution. Simulation results underscore the accuracy and optimal performance of the proposed method compared to K-Nearest Neighbor, Gradient Boosting Regression, and Linear Regression methods.