<p>Seawater salinity is a critical parameter governing the dynamics of harmful algal blooms (HABs), making accurate short-term forecasting essential for marine ecosystem management. While deep learning has advanced significantly, existing models for short-term salinity forecasting face challenges in accuracy and generalization. We present an artificial intelligence framework based on the inverted Transformer (iTransformer) architecture for short-term seawater salinity forecasting. The model incorporates a multi-scale embedding module to capture temporal patterns across different scales and employs a sparse attention mechanism within the Transformer encoder to reduce computational complexity while preserving long-range temporal dependencies. The Huber loss function enhances model robustness against anomalous observations. We evaluated the model against eight state-of-the-art time series forecasting approaches using four in situ salinity datasets from British Columbia, Canada. Results show that the iTransformer-based model consistently achieves superior predictive accuracy across all datasets, indicating its potential for operational deployment in coastal monitoring and HABs early warning systems.</p>