Non-Autoregressive Machine Learning for Spring Discharge Forecasting: A Bias-Corrected CNN–LSTM Approach

Abstract

Sustainable water resource management in semi-arid Mediterranean regions is critically dependent on reliable spring discharge forecasting and on quantifying the long-term impacts of climate change on groundwater resources. This study develops and evaluates a suite of non-autoregressive machine learning architectures, MLP, CNN, LSTM, and a hybrid CNN–LSTM, for predicting the monthly discharge of major springs supplying Palermo, Italy. Unlike autoregressive approaches, which are susceptible to error accumulation in multi-step projections and cannot be applied under projected climate conditions, the proposed models rely exclusively on exogenous meteorological inputs, making them inherently suited for operational forecasting and architecturally compatible with long-term climate impact assessment. The CNN–LSTM architecture achieves the best performance among all non-autoregressive configurations, approaching the idealized autoregressive benchmark while avoiding its operational limitations. In the operational phase, the calibrated CNN–LSTM model is driven by bias-corrected ECMWF SEAS5 seasonal forecasts, maintaining useful predictive performance up to six months ahead for RIS, SCI, and the aggregated system (SYS), with more limited skill for GAB at longer lead times. Diebold–Mariano tests confirm that this advantage is statistically significant for two of the three springs, whereas for the third, whose dynamics are dominated by internal aquifer memory, no non-autoregressive architecture prevails. This framework not only enables retrospective analysis of spring discharge behavior but is also, thanks to its non-autoregressive design, structurally suited for future coupling with climate projections, offering water authorities an effective solution for proactive and sustainable management of this crucial Mediterranean water resource.