Balancing accuracy and efficiency in multi-step building-level electric load forecasting: Deep learning vs. classical time-series models

Accurate short-term electric load forecasting is vital for maintaining the stability and efficiency of modern power systems, particularly at the building level, where consumption patterns are highly dynamic and weather-dependent. Most existing methods focus on system-level aggregation, provide only single-step forecasts, or fail to balance accuracy, robustness, and computational efficiency under volatile conditions. To bridge these gaps, this study introduces novel hybrid models, BOA-NARX and BOA-LSTM, which integrate Bayesian Optimization Algorithms (BOA) with advanced recurrent neural networks: non-linear autoregressive networks with exogenous inputs (NARX) and long short-term memory (LSTM), respectively. This study's novelty is encapsulated in the following contributions: (i) compares conventional SARIMAX (seasonal autoregressive integrated moving average with exogenous inputs) vs. deep learning (LSTM, NARX) for multi-step day-ahead building-level load forecasting; (ii) automatic hyperparameter tuning via BOA to enhance performance without manual intervention; (iii) evaluation of model robustness under varying levels of input uncertainty, such as noisy weather forecasts; (iv) an in-depth analysis of seasonal load volatility and its influence on forecast accuracy; and (v) empirical guidance for adaptive model selection based on load volatility, suggesting optimal scenarios for deploying either deep learning or classical statistical models. Experimental results demonstrate that BOA-LSTM consistently achieves the highest forecasting accuracy (R² > 0.99) across all seasons, while SARIMAX remained competitive under low-volatility conditions with minimal computational cost. These findings support the development of adaptive, intelligent forecasting frameworks that are both accurate and resilient, enabling smarter energy management in buildings and more responsive power systems.