Multi-step ahead streamflow forecasting method using Embedding Multi-Layer Perceptron

Three flood-prone watersheds in Iowa, U.S.: the Clear Creek Watershed, the Upper Wapsipinicon River Watershed, and the Iowa River Watershed. Reliable and computationally efficient multi-step ahead streamflow forecasting remains a key challenge in water resources management. While data-driven hydrological models show considerable potential, their performance often degrades with increasing forecast horizons, and they typically require substantial computational resources. To address these limitations, this study proposes a physically plausible yet lightweight deep learning framework, the Embedding Multi-Layer Perceptron (E-MLP), to enhance reliability and efficiency in multi-step ahead streamflow forecasting. The proposed model combined a flux-pattern learning module capturing rapid hydrological responses with a learnable latent factor that modulates historical information for future prediction. For evaluation, four benchmark models (Persistent, Ridge regression, Long Short-Term Memory, and Multi-Layer Perceptron models) were implemented. Comparative experiments were conducted using hourly rainfall-streamflow data from the study region and forecast horizons of up to 12 h. The E-MLP model consistently outperformed the benchmarks, effectively mitigating forecast-horizon-dependent accuracy loss while maintaining high computational efficiency. Component-wise analyses indicated that the separating flux-pattern learning from historical information modulation contributes to the improved predictive skill. Cross-basin analyses suggested that E-MLP organizes and modulates historical information in a manner consistent with known scale-dependent hydrological responses within the study region, potentially supporting reliable streamflow forecasting across watershed scales. • A lightweight, physically plausible E-MLP model is proposed for streamflow forecasting. • Multi-watershed tests show robust, efficient, and interpretable long-lead predictions.