Work

Long-Term Peak Load Forecasting with CNN Via Two-Stage Random–Bayesian Hyperparameter Optimization

Year: 2025
Type: article
Abstract: This study addresses the challenge of long-term peak load forecasting in power systems by enhancing Convolutional Neural Networks (CNN) with an efficient two-stage hyperparameter optimization strategy. The research contribution is a sequential Random–Bayesian approach that first explores the hyperparameter space broadly using Random Search and then refines the search with Bayesian Optimization, improving prediction accuracy while controlling computational cost. Experiments on daily peak load data from New South Wales (Australia), spanning 2015–2022, demonstrate that the proposed two-stage strategy achieves the lowest mean absolute error (MAE = 695.57), mean absolute percentage error (MAPE = 7.39%), and competitive root mean square error (RMSE = 900.68) compared to single-stage Random Search and Bayesian Optimization. Although the two-stage method requires higher training time than Random Search, it offers a balanced trade-off between accuracy and computational efficiency. These findings provide a practical reference for designing robust hyperparameter optimization strategies in CNN-based long-term load forecasting, thereby supporting more informed decision-making for grid reliability and capacity planning.
Language: English
Open Access status: diamond
OpenAlex ID: W7160911374
FWCI: 0.174
Cites:
Cited by: 1
Related to: 10