Optimizing N-BEATS for Short-Term Load Forecasting Via Random Search–TPE and Genetic Algorithm

Accurate short-term electricity load forecasting is crucial for secure and economical power system operation; however, deep forecasting models can degrade markedly when hyperparameters are poorly chosen. This paper presents a systematic and reproducible hyperparameter-optimization framework for the N-BEATS (Neural Basis Expansion Analysis for Time Series) model to improve one-day-ahead load forecasting under a fixed experimental setup. A unified tuning protocol-using the same objective, search space, and evaluation budget-optimizes key N-BEATS hyperparameters (stack_types, n_blocks, mlp_units, and learning_rate). Three strategies are compared fairly: Random Search, Tree-structured Parzen Estimator (TPE) Bayesian optimization, and a Genetic Algorithm (GA), each with 100 evaluations (100 trials). Experiments use two Australian National Electricity Market regional datasets (New South Wales and Queensland), each with 122,735 half-hourly demand records; the most recent 28 days are used for training and the final day for testing, with a 7-day input window (336 points) forecasting 48 steps (24 hours) ahead (max_steps = 1000). Performance is assessed using MSE, RMSE, MAE, and MAPE. All tuned configurations outperform the default N-BEATS baseline. Using MAPE as the optimization objective, TPE performs best in both regions. In NSW, MAPE decreases from 2.56% to 1.29% (49.6% reduction), and MAE decreases from 180.0 MW to 88.7 MW (50.7% reduction). In QLD, MAPE decreases from 2.80% to 1.79% (36.1% reduction), while also yielding the lowest MSE/RMSE/MAE. These results confirm the value of standardized hyperparameter tuning for N-BEATS and suggest that the most effective strategy can be region- and metric-dependent.