BiLSTM with Asymmetric Huber Loss for 24 h PM2.5 Forecasting and Early Exceedance Detection at an Industrial Monitoring Station in Bogotá, Colombia

Fine particulate matter (PM2.5) pollution in Bogotá follows a strongly asymmetric temporal pattern: at the Puente Aranda industrial-corridor monitoring station, 98.7% of days whose daily mean exceeds the WHO 2021 guideline of 15 μg/m3 contain at least one hourly concentration above 25 μg/m3, with a median peak-to-mean ratio of 1.90. Conventional forecasting models, trained to minimise average error, systematically underpredict these critical episodes. To address this structural bias, we trained a Bidirectional Long Short-Term Memory (BiLSTM) network for 24 h PM2.5 forecasting using three complementary mechanisms: an asymmetric Huber loss that penalises underprediction α-fold more than overprediction, sample weighting that amplifies gradients from exceedance windows, and Bayesian hyperparameter optimisation (Optuna-TPE, 100 trials). Evaluating four model configurations on hourly records from 2022 to 2025 (n = 31,961), we identify an inherent regression–detection tension: among the α values examined (α∈{1.5,2.5,2.80}), no single value simultaneously minimised global MAE and peak MAE. Bayesian optimisation (Optuna-TPE, 100 trials) selected α=2.80 and whigh=10, achieving the nominally highest sequence-level detection quality (AP = 0.724, 95% bootstrap CI [0.582, 0.821]) and the lowest MAEpeaks (10.64 μg/m3), but at the cost of near-zero global regression (R2≈0). BiLSTM-A1.5 (α=1.5) constitutes the best overall compromise (AP = 0.718, CI [0.575, 0.819]; RMSE = 6.65 μg/m3; R2=0.372; sequence F1 = 0.674); the AP difference (Δ=0.006) is within overlapping bootstrap confidence intervals. All models localise the daily peak within 6 h in at least 88% of exceedance sequences. These results support reframing early-warning system evaluation around exceedance detection rather than regression error.