Enhancing Building Energy Forecasting: A Comparative Analysis of LSTM, GRU, and RNN Models for Active Power, Reactive Power, and Power Factor Prediction

The increasing need for efficient energy management in buildings drives the importance of accurate load forecasting. This study focuses on predicting critical yet underexplored parameters, reactive power and power factor, essential for optimizing energy efficiency. A systematic evaluation of classical deep learning models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Recurrent Neural Network (RNN) is conducted to address this gap. While these architectures are foundational in temporal forecasting, their application here leverages unique strengths: LSTM's memory cells for long-term dependencies, GRU's gating mechanisms for computational efficiency, and RNN's sequential data processing. By adapting them to multi-parameter energy forecasting (active power, apparent power, and power factor), this work demonstrates their relevance in modern systems requiring multivariate time-series analysis.Arigorous feature selection process identifies optimal predictors for each parameter, and models are optimized for multi-horizon predictions to address operational and planning needs. The case study utilizes high-resolution smart meter data from the Universidad de Las Américas (UDLA) in Ecuador, emphasizing practical implementation in real-world educational infrastructures. The findings provide actionable insights for enhancing sustainability in buildings, particularly in institutions managing dynamic consumption patterns and grid demands.