Accurate long-horizon air-quality forecasting becomes difficult when historical observations are missing or irregularly sampled because reconstruction errors can propagate into downstream predictions. In this work, we propose the TILSTM method, a task-aligned hybrid architecture that integrates a Transformer-based imputation module with an LSTM forecaster designed to jointly enforce a causal horizon boundary that restricts imputation strictly to the historical look-back window, an observed-preserving merge that leaves measured values unchanged, and a time-aware decay gate applied selectively to imputed positions. The model is trained end-to-end using a combined forecasting loss and a self-supervised imputation loss computed on artificially masked observed entries. We evaluate TILSTM on hourly PM10 forecasting from 21 monitoring stations in Slovenia across three forecasting horizons and three missingness regimes. Among the compared methods, TILSTM shows the clearest and most consistent gains at the 24 h horizon, while at medium horizons, the relative ranking becomes more dependent on the missingness regime. In pooled error summaries, TILSTM achieves the lowest MAE and RMSE at the 168 h horizon under the real and nearₒrigin missingness regimes, while the overall results indicate that no single method is uniformly best across all long-horizon settings.
Vrbančič et al. (Fri,) studied this question.