Bayesian-Optimized MLP-LSTM-CNN for Multi-Year Day-Ahead Electric Load Forecasting

Accurate long-term electric load forecasting—multi-year, day-ahead peak-load prediction—is critical for planning, operations, and policy. While traditional statistical and shallow machine-learning methods often struggle with nonlinear and multi-scale temporal patterns, deep learning offers promising alternatives. This study conducts a systematic, controlled comparison of three architectures—Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)—within a unified Bayesian hyperparameter optimization protocol using daily peak-load data from the New South Wales (NSW) electricity market, 2015–2021, with a 365-day look-back window. Under identical data splits, objective, and search procedures, CNN delivers the best accuracy across all metrics (MAE = 699, MSE = 791,838, RMSE = 890, MAPE = 7.53%), MLP performs slightly worse, and LSTM yields the most significant errors alongside the most extended runtime. These results indicate that, under consistent tuning and a one-year context window, CNN captures local variations more effectively than the recurrent alternative in this setting. The research contribution of this study is a fair, empirical benchmark of widely used deep models (MLP, CNN, LSTM) for multi-year, day-ahead peak-load forecasting under a single Bayesian optimization framework, offering practical guidance for model selection. Reproducibility is facilitated by fixed random seeds and comprehensive configuration/trace logging. Limitations include an intentionally univariate design (no exogenous variables), a focus on learned architectures rather than naïve baselines, and the absence of uncertainty quantification; future work will extend to multivariate inputs (e.g., weather and calendar effects), hybrid CNN–LSTM and Transformer-based models, and broader baseline and robustness evaluations.