FPQ-iTransformer: A novel iTransformer-based model for accurate seawater salinity prediction

Abstract

Accurate seawater salinity prediction is crucial for marine ecosystem management yet challenged by climate-driven dynamic fluctuations. We develop an enhanced iTransformer model FPQ-iTransformer incorporating three novel components: ProbSparse self-attention in the encoder for efficient temporal pattern extraction, a hybrid decoder combining multivariate attention with sparsity mechanisms for spatiotemporal dependency capture, and quantile loss integration for noise-robust prediction. Validated on British Columbia’s mariculture monitoring data, our framework demonstrates superior performance over eight state-of-the-art models (CNN, LSTM, Transformer, etc.), achieving 12.7% MAE reduction and 15.4% $R^2$ improvement compared to the best baseline. The model’s prediction accuracy shows increasing advantages with extended time horizons, particularly in capturing abrupt salinity variations caused by extreme weather events. This advancement provides a reliable computational tool for real-time aquaculture decision-making and coastal environmental protection, with potential applicability to other hydrodynamic parameter predictions in climate-sensitive regions.