Hybrid Bayesian–Hyperband Optimization of a One-Dimensional Convolutional Neural Network for Short-Term Load Forecasting

Short-term load forecasting (STLF) plays a crucial role in power system operation and electricity market planning. However, the forecasting accuracy of convolutional neural network (CNN) models strongly depends on properly tuned hyperparameters such as filter numbers, batch size, and loss function, while exhaustive tuning is computationally expensive. This paper proposes a one-dimensional CNN (1D-CNN) for STLF whose key hyperparameters are optimized by a hybrid Bayesian Optimization–Hyperband framework (BO-TPE-HB). The framework combines the Tree-structured Parzen Estimator (TPE), which guides the search towards promising regions of the hyperparameter space, with Hyperband’s multi-fidelity early-stopping strategy to terminate weak configurations early and save computation. The proposed approach is evaluated using half-hourly electricity load data from the Australian Energy Market Operator (AEMO) for New South Wales (NSW) and Victoria (VIC) from 2009 to 2014. Using a rolling-origin time-series cross-validation scheme on the training set and a chronologically separated hold-out test set, BO-TPE-HB explores 100 candidate configurations of a 4-layer 1D-CNN with a sliding window of 48 time steps. Model performance is assessed using MAE, MSE, RMSE, MAPE, and total hyperparameter optimization runtime. Experimental results show that the BO-TPE-HB–tuned CNN achieves test MAPE of 1.350% (MAE 105.98 MW, RMSE 143.71 MW) for NSW and 1.615% (MAE 89.13 MW, RMSE 124.25 MW) for VIC, outperforming Random Search, standalone TPE, and standalone Hyperband in terms of both prediction accuracy and computational efficiency, while requiring substantially less tuning time than plain Bayesian Optimization. These findings highlight that combining probabilistic search with resource-efficient early stopping provides a practical and reproducible way to enhance CNN-based STLF, and can be extended to multivariate inputs or other deep learning forecasting architectures in future work.