Investigating The Impact of Hyperparameters on N-BEATS for Load Forecasting

This study investigates how key architectural and training hyperparameters influence the forecasting accuracy of the N-BEATS deep learning model for short-term load forecasting. Rather than proposing a new automated optimization algorithm, the study addresses the need for a controlled, interpretable sensitivity analysis of N-BEATS within a fixed experimental setting. The research contribution is to identify which hyperparameter groups most strongly affect forecasting performance and to determine practical configuration ranges for N-BEATS-based load forecasting. Experiments are conducted on the Tasmania daily peak-load dataset from AEMO, which contains 2,120 daily observations. A fixed split is used, with 364 days for training and a 7-day forecasting horizon for testing. The evaluation is carried out in two stages: 45 architectural configurations are first examined by varying stack variant, number of stacks, and number of blocks per stack, and then 125 training-hyperparameter configurations are evaluated by varying maximum training steps, learning rate, and MLP width. The results show that N-BEATS performance is strongly affected by both architectural and training hyperparameters. Within the evaluated search space, the best-observed configuration achieves a MAPE of 3.1300% and an RMSE of 51.35, compared with 4.2839% and 64.32 for the default configuration, corresponding to a 26.9% reduction in MAPE. Moderate block depths perform better than overly shallow or deeper settings. In addition, max_steps = 500, learning rates from 1x10^4 to 5x10^4, and an MLP width of [512, 512] provide the most favorable accuracy–stability trade-off. These findings provide practical guidance for configuring N-BEATS in load-forecasting applications.