Smart Time Series Energy Forecasting in Buildings: Principal Variate Selection Approach for Long Short-Term Memory Models

<p dir="ltr">Accurate building load forecasting is critical for improving energy efficiency in smart construction systems and data-driven facility management. While deep learning models such as long short-term memory (LSTM) have shown promising performance in capturing complex temporal patterns, their effectiveness relies heavily on the quality and relevance of input features. In this research, we propose a Principal Variate Selection (PVS) forecasting pipeline that integrates Bayesian Optimization to identify the optimal feature subset for multivariate LSTM models. Although multivariate models generally outperform univariate ones, incorporating all available features can lead to exponential increases in computational costs and degrade model performance. Our approach ensures superior predictive accuracy compared to univariate baselines while reducing the dimension of inputs, thereby improving both accuracy and training efficiency. The pipeline supports deployment in scenarios where data acquisition is costly or constrained, such as early-stage smart building projects or resource-limited environments. This work offers a generalizable method for improving prediction quality while optimizing building load data col lection, contributing to the actionable insights for achieving more precise and reliable load forecasts targeting specific energy management needs.</p>