Environmental monitoring systems face persistent challenges with missing data, particularly in air quality networks where sensor failures, power outages, and maintenance activities frequently create gaps in temporal measurements. Traditional imputation methods struggle with the sophisticated temporal dynamics and complex relationships inherent in environmental data. This paper introduces a novel Temporal Variational Autoencoder (VAE) approach that addresses the fundamental limitation of standard VAEs in handling missing data through learnable missing embeddings. Instead of using arbitrary placeholder values for missing data, our method learns optimal representations for missing values during training, enabling VAEs to process partial observations naturally. We implement variable-type-specific learnable embeddings that account for different characteristics of meteorological versus pollutant variables, combined with a temporal VAE architecture optimized for environmental time series with multiple temporal scales. Comprehensive evaluation on EPA air quality datasets demonstrates substantial improvements across various missing scenarios, with 15–20% RMSE reductions compared to standard VAE approaches and 25–35% improvements over traditional statistical methods. The proposed method maintains computational efficiency suitable for deployment in operational environmental monitoring networks while providing uncertainty quantification crucial for downstream analysis and decision-making.