This study explores a hybrid AI framework for streamflow forecasting that integrates physically based hydrological modeling, bias correction, and deep learning. HEC-HMS simulations generate synthetic discharge, which a machine learning-based bias correction model adjusts for irrigation-induced discrepancies—improving the Nash–Sutcliffe Efficiency (NSE) from 0.55 to 0.84, the Kling–Gupta Efficiency (KGE) from 0.67 to 0.89, and reducing the RMSE from 1.084 to 0.301 m³/s. The corrected discharge is used as input to a Temporal Fusion Transformer (TFT) trained on hourly meteorological data to predict streamflow at 24-, 48-, and 72-h horizons. In a semi-arid, irrigated basin in Northern Greece, the TFT achieves NSEs of 0.84, 0.78, and 0.71 and RMSEs of 0.301, 0.743, and 0.980 m³/s, respectively. Probabilistic forecasts deliver uncertainty bounds with coverage near nominal levels. In addition, the model’s built-in interpretability reveals temporal and meteorological influences—such as precipitation—that enhance predictive performance. This framework demonstrates the synergistic benefits of combining physically based modeling with state-of-the-art deep learning to support robust, multi-horizon forecasts in irrigation-influenced, data-scarce environments.

Time series forecasting plays a critical role across numerous domains such as finance, energy, and healthcare. While traditional statistical models have long been employed for this task, recent advancements in deep learning have led to a new generation of state-of-the-art (SotA) models that offer improved accuracy and flexibility. However, there remains a gap in understanding how these forecasting models perform under different forecasting scenarios, especially when incorporating external variables. This paper presents a comprehensive review and empirical evaluation of seven leading deep learning models for time series forecasting. We introduce a novel dataset that combines energy consumption and weather data from 24 European countries, allowing us to benchmark model performance across various forecasting horizons, granularities, and variable types. Our findings offer practical insights into model strengths and limitations, guiding future applications and research in time series forecasting.