Machine learning-based multivariate time-series forecasting for environmental monitoring in pig houses under high stochasticity

Ammonia gas concentration in intensive swine production pose serious environmental and health challenges, particularly in naturally ventilated systems. This study presents an application-oriented comparative evaluation of eight state-of-the-art forecasting models to predict NH 3 concentrations in a commercial pig house. First a factorial design assessed the impact of covariates, resolution time, window lengths, and pre-processing. Results identified data preprocessing as the primary driver of model accuracy, where causal non-centered smoothing significantly outperformed aggressive outlier removal, and the exclusion of non-essential covariates for predictions. Among the architectures, TiDE and D-Linear excelled in short-term forecasting. Targeted optimization of the TiDE model using only past covariates achieved an R 2 of 0.895 ± 0.005 and a Mean Absolute Error of 0.150 ± 0.012 for a 30-minute horizon, representing an improvement in variance explanation over the unoptimized baseline. For longer 60-minute horizons, models relying solely on past covariates showed diminishing returns; the optimized linear model (D-Linear) improved its R 2 to 0.507 ± 0.005, while the deep TiDE architecture plateaued at an R 2 of 0.583 ± 0.062. However, integrating future weather covariates proved essential for the Temporal Fusion Transformer (TFT), which utilized these inputs to stabilize multi-horizon forecasts, achieving a highly stable R 2 of 0.88 ± 0 at 30 min and maintaining a robust R 2 of 0.665 ± 0.015 at the challenging 60-minute horizon. These findings suggest a hybrid Precision Livestock Farming framework: employing optimized lightweight models like TiDE for immediate, real-time ventilation control, while reserving attention-based architectures enriched with meteorological forecasts for extended operational planning.