To improve coastal adaptation and management, it is critical to better understand and predict the characteristics of sea levels. Here, we explore the capabilities of artificial intelligence, from four deep learning methods to predict the surge component of sea-level variability based on local atmospheric conditions. We use an Artificial Neural Networks, Convolutional Neural Network, Long Short-Term Memory layer (LSTM) and a combination of the latter two (ConvLSTM), to construct ensembles of Neural Network (NN) models at 736 tide stations globally. The NN models show similar patterns of performance, with much higher skill in the mid-latitudes. Using our global model settings, the LSTM generally outperforms the other NN models. Furthermore, for 15 stations we assess the influence of adding complexity more predictor variables. This generally improves model performance but leads to substantial increases in computation time. The improvement in performance remains insufficient to fully capture observed dynamics in some regions. For example, in the tropics only modelling surges is insufficient to capture intra-annual sea level variability. While we focus on minimising mean absolute error for the full time series, the NN models presented here could be adapted for use in forecasting extreme sea levels or emergency response.

Reliable sub-seasonal forecasts of high summer temperatures would be very valuable for society. Although state-of-the-art numerical weather prediction (NWP) models have become much better in representing the relevant sources of predictability like land- and sea-surface states, the sub-seasonal potential is not fully realized. Complexities arise because drivers depend on the state of other drivers and on interactions over multiple time-scales. This study applies statistical modeling to ERA5 reanalysis data, and explores how nine potential drivers, interacting on eight time-scales, contribute to the sub-seasonal predictability of high summer temperatures in western and central Europe. Features and target temperatures are extracted with two variations of hierarchical clustering, and are fitted with a machine learning (ML) model based on Random Forests. Explainable AI methods show that the ML model agrees with physical understanding. Verification of the forecasts reveals that a large part of predictability comes from climate change, but that reliable and valuable sub-seasonal forecasts are possible in certain windows, like forecasting monthly warm anomalies with a lead time of 15 days. Contributions of each driver confirm that there is a transfer of predictability from the land- and sea-surface state to the atmosphere. The involved time-scales depend on lead time and the forecast target. The explainable AI methods also reveal surprising driving features in sea-surface temperature and 850 hPa temperature, and rank the contribution of snow-cover above that of sea-ice. Overall, this study demonstrates that complex statistical models, when made explainable, can complement research with NWP models, by diagnosing drivers that need further understanding and a correct numerical representation, for better future forecasts.

The succession of European surface weather patterns has limited predictability because disturbances quickly transfer to the large‐scale flow. Some aggregated statistics, however, such as the average temperature exceeding a threshold, can have extended predictability when adequate spatial scales, temporal scales and thresholds are chosen. This study benchmarks how the forecast skill horizon of probabilistic 2‐m temperature forecasts from the subseasonal forecast system of the European Centre for Medium‐Range Weather Forecasts (ECMWF) evolves with varying scales and thresholds. We apply temporal aggregation by rolling‐window averaging and spatial aggregation by hierarchical clustering. We verify 20 years of re‐forecasts against the E‐OBS dataset and find that European predictability extends at maximum into the fourth week. Simple aggregation and standard statistical post‐processing extend the forecast skill horizon with two and three skilful days on average, respectively. The intuitive notion that higher levels of aggregation capture large‐scale and low‐frequency variability and can therefore tap into extended predictability holds in many cases. However, we show that the effect can be saturated and that there exist regional optimums beyond which extra aggregation reduces the forecast skill horizon. We expect such windows of predictability to result from specific physical mechanisms that only modulate and extend predictability locally. To optimize subseasonal forecasts for Europe, aggregation should thus be limited in certain cases.