Accurate assessment of lake water balance is vital for sustainable management in data-scarce regions. Lake Tana, Ethiopia’s largest freshwater lake and source of the Blue Nile, faces growing uncertainty due to limited streamflow monitoring, land-use change, and climate variability. This study developed a hybrid framework integrating the DeepGR4J rainfall-runoff model with regression-based regionalization to simulate inflows and outflows from 2000-2020. DeepGR4J combines the GR4J conceptual model with LSTM networks to correct residual errors, while ungauged catchments were represented using physiographic and remote sensing descriptors. Precipitation inputs were derived from bias-corrected CHIRPS data, validated against 20 stations, and evapotranspiration was estimated using the Hargreaves method. Results show mean annual inflow of 12 km³ yr⁻¹ and outflow of 12.87 km³ yr⁻¹, yielding a net storage decline of -0.26 km³ yr⁻¹ and residual losses of 0.35 km³ yr⁻¹, likely reflecting uncertainties and groundwater seepage. Model performance was excellent (R² = 0.94; NSE = 0.94; RMSE = 0.18 m), confirming reliability in reproducing lake-level dynamics. This transferable hybrid framework offers significant insights for water balance assessment and adaptive management in data-limited basins.