Short-Term Regional Precipitation Forecasting over Dublin Using ERA5 Reanalysis: Robustness, Probabilistic Reliability, and Predictor Attribution

Predicting rain is one of the most stubborn challenges in meteorology. Unlike the steady rise of the sun or the predictable pull of the tides, rainfall is erratic, driven by a chaotic mix of air pressure, wind, and moisture that can shift in minutes. This unpredictability is especially acute in places like Ireland, where the Atlantic Ocean constantly pushes weather systems against the land, creating conditions that change rapidly. For cities like Dublin, getting the forecast right for the next few hours is not just a matter of convenience; it is a safety issue. Accurate short-term predictions help manage urban flooding, keep roads safe for drivers, and allow water managers to plan for the future. For decades, scientists have relied on massive, physics-based computer models to simulate the atmosphere, but these systems are heavy and slow. In recent years, a new approach has emerged: using artificial intelligence to learn patterns from historical weather data, hoping to spot the signals of an approaching storm faster and more efficiently than traditional methods.

A team of researchers at University College Dublin and Trinity College Dublin set out to test how well these modern machine-learning tools actually work for short-term rain forecasting in their own backyard. They focused on a specific question: can a computer learn to predict the average rainfall over the Dublin region six hours into the future, and if so, how reliable is that prediction when the weather turns severe? To answer this, they did not just build one model and hope for the best. Instead, they gathered a massive library of historical weather data from the European Centre for Medium-Range Weather Forecasts, covering every six-hour interval from 2004 to 2021. They fed this data into several different types of forecasting systems, ranging from simple statistical benchmarks to sophisticated deep-learning networks that can remember past weather patterns and weigh them against current conditions.

The researchers treated this like a rigorous scientific trial rather than a simple demonstration. They ran their experiments multiple times with different starting conditions to ensure the results were not just a lucky fluke. They also looked beyond the average performance. In weather forecasting, a model that is good at predicting a light drizzle might fail miserably when a heavy downpour hits. The team specifically checked how each model handled wet days and, more importantly, the most intense rainfall events. They also asked the models to provide a sense of uncertainty, not just a single number, so that forecasters would know how confident they could be in the prediction. Finally, they peered inside the "black box" of the artificial intelligence to see which weather factors the computer was actually paying attention to, distinguishing between variables that were merely correlated with rain and those that truly helped the model make better predictions.

The results revealed a landscape of performance that was more nuanced than a simple ranking of "best" versus "worst." When looking at the overall picture across all types of weather, a specific type of deep-learning network called a BiLSTM with an attention mechanism performed slightly better than its competitors. This model, which can look back at a sequence of past weather data and focus on the most relevant moments, achieved a very low average error of about 0.24 millimeters. However, the difference between this top performer and a powerful tree-based algorithm called XGBoost was so small that, statistically, they were essentially tied for the general test set. The researchers found that for the vast majority of days, including those with some rain, these two very different approaches produced nearly identical results.

The story changed, however, when the weather turned severe. As the researchers isolated the heaviest rainfall events—those falling in the top five percent of intensity—the performance of the models diverged. In these difficult conditions, the XGBoost algorithm actually outperformed the deep-learning network, producing more accurate forecasts for the heaviest downpours. This finding suggests that while complex neural networks are excellent at learning general patterns, simpler, tree-based methods might be better at handling the rare, extreme outliers that are most critical for flood warnings. Furthermore, the ability of all models to detect rain events dropped sharply as the intensity increased. While they could reliably predict that it would rain at all, their skill at pinpointing exactly how hard it would fall diminished rapidly as the threshold for "heavy" rain rose.

The study also examined the reliability of the models' confidence. The probabilistic versions of the deep-learning models, which were designed to give a range of possible outcomes rather than a single guess, performed well on average. Over the entire test period, their predicted ranges captured the actual rainfall about 95 percent of the time, matching their intended confidence level. But this reliability broke down when it mattered most. For the heaviest rainfall events, the models became overconfident; their predicted ranges were too narrow, and they missed the actual heavy rain far more often than they should have. Additionally, because the models assumed rainfall followed a standard bell-curve distribution, they occasionally predicted negative amounts of rain—a physical impossibility—highlighting a fundamental mismatch between the mathematical assumptions and the real-world behavior of water falling from the sky.

To understand why the models made the choices they did, the researchers used a technique to trace the influence of each weather variable. They discovered that the computer's attention was drawn most strongly to total cloud cover, the direction of the wind, and surface pressure. Interestingly, the variable with the strongest raw statistical link to rain—maximum precipitation rate—was not the most important driver for the model's decisions once other factors were considered. This is because that variable is so closely tied to others that it offers little new information. In contrast, total cloud cover, which had a weaker raw connection to rain, provided unique and stable predictive value that the model relied on heavily. This distinction showed that simply looking at which numbers move together is not enough; understanding how a model uses those numbers to make a decision requires looking deeper.

Ultimately, this work underscores that there is no single "perfect" model for forecasting rain. The best approach depends on what you are trying to predict. For general, short-term outlooks, a compact deep-learning network and a robust tree-based algorithm perform equally well. But for the specific, high-stakes task of predicting severe storms, the tree-based method showed a distinct edge. The study also serves as a reminder that a model's overall accuracy can hide significant weaknesses in extreme conditions. A system that looks reliable on average may fail to warn of the most dangerous weather. By combining repeated testing, a focus on extreme events, and a careful look at how the models think, the researchers provided a clearer, more honest picture of what artificial intelligence can and cannot do for the people of Dublin when the clouds gather.