Impact of Input Feature Scenarios on Metaheuristic-Optimized LSTM and GRU Networks Applied to Load Forecasting for Building Energy Management

Accurate short-term load forecasting is a key enabler of energy management in smart buildings, yet the performance of recurrent neural network models is highly sensitive to hyperparameter configuration---a challenge that manual tuning addresses poorly. This paper proposes a metaheuristic-based hyperparameter optimization (HPO) framework that applies Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) to automatically configure Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures for one-hour-ahead building load forecasting. The framework is evaluated across four input scenarios of increasing feature richness using 18 months of real measurements from the CU-BEMS dataset, a seven-story academic building in Bangkok, Thailand. Results show that the inclusion of outdoor temperature as an exogenous variable (Scenario S3) yields the best overall performance, with LSTM-PSO achieving RMSE = 20.15 kW and \((R^2 = 0.9876)\)---a 19% reduction in RMSE relative to the univariate baseline. HPO provides modest but consistent improvements over fixed-parameter baselines (up to 3.9% in Scenario S3), while GA and PSO produce statistically equivalent results across all configurations. GRU models exhibit greater stability across runs than LSTM models, particularly under high-dimensional input settings. The proposed framework requires only standard Building Energy Management System (BEMS) sensor outputs and offers a practical, automated pathway for deploying accurate load forecasting in commercial and institutional buildings.