Univariate vs. multivariate long-short term memory for daily rainfall forecasting at a coastal station in New South Wales

Rainfall forecasting plays a pivotal role in environmental planning and disaster management. This pilot study evaluates two Long Short-Term Memory (LSTM) architectures for next-day rainfall prediction at a single Australian station (Albion Park, NSW). The first model is a univariate LSTM (U-LSTM) using a seven-day window of antecedent rainfall to predict the following day. The second is a multivariate LSTM (M-LSTM), which augments the same temporal window with co-observed meteorological drivers (temperature, humidity, wind speed, solar radiation, and evapotranspiration). Both models are implemented in PyTorch (Python based machine learning library), trained with Adam, and evaluated using standard hydrological metrics. The M-LSTM achieves an RMSE of 0.987 mm and MAE of 0.378 mm on the test data set, with weak correlation (r = 0.176) and negative Nash–Sutcliffe Efficiency (NSE = −0.077), indicating limited skill in reproducing day-to-day amounts. Visual diagnostics reveal systematic underestimation and variance compression. While the multivariate setting provides small but consistent improvements over the univariate baseline, both models struggle with zero-inflation and extremes. The paper concludes with a prioritized roadmap—data, architecture, loss design, and evaluation—to guide future extensions (e.g., multi-station learning, probabilistic outputs, hybrid physical–statistical models) suited to Australian conditions.