HoST: integrating Heuristic knowledge with attention-based LSTM networks for Thunderstorm prediction.

Interpretable Daily Inflow Forecasting Model via Cross-Attention Hybrid Neural Networks Integrated with SHAP Analysis

BiasCast: learning and adjusting real time biases from meteorological forecasts to enhance runoff predictions

Abstract. The use of deep learning models in hydrology is becoming an ever more prevalent application in operational flood forecasting. Such operational systems face performance degradation when transitioning from high quality reanalysis to meteorological forecast data with lower accuracy. This study investigates training strategies and Long Short-Term Memory network architectures to mitigate meteorological forecast-induced bias in maximum daily discharge predictions using the Extended LamaH- CE dataset and a subset of 451 basins. We systematically evaluated cross-domain generalization, transfer learning approaches, Encoder–Decoder LSTMs, Sequential Forecast LSTMs, and the role of input embeddings and integrating past discharge observations. The results show that domain shifts between reanalysis and forecast data lead to substantial skill loss, with median Nash–Sutcliffe Efficiency decreasing from 0.58 to 0.33. Among the tested strategies, the Sequential Forecast LSTM demonstrated the most stable improvements, achieving a median NSE of 0.63. Integrating recent discharge observations further enhanced performance, raising median NSE to 0.71 and surpassing even the reanalysis-driven baseline. In contrast, integrating archived forecasts or using more complex input embeddings did not yield consistent benefits and in some cases degraded model stability. Basin-level analysis reveals that forecast skill improvements compared to our baseline are not uniformly distributed across catchment types: the largest gains are concentrated in arid and precipitation-limited catchments, while alpine and snow-dominated catchments, despite experiencing the largest meteorological domain shift, show smaller improvements. This is likely because the LSTM cell state retains strong seasonal signals and thereby compensates for forecast input bias through its long-term memory mechanism, particularly in alpine and snow-dominated catchments where strong seasonal cycles dominate the hydrological response. These findings highlight the value of training strategies that allow models to directly learn bias correction during forecast transitions, emphasize the operational potential of combining sequential processing with near real-time discharge observations and identify physiographic catchment characteristics as key modulators of forecast skill improvement across diverse hydroclimatic settings.

Deep Learning Based Climate Forecasting Model Using LSTM (Long Short-Term Memory) of Kathmandu Valley

Reliable temperature and precipitation forecasting is more important than ever for communities and policymakers preparing for the challenges ahead for effective adaptation and mitigation strategies in the Kathmandu Valley, which is located within a landlocked mountain valley. A deep learning-based climate forecasting model for the area is proposed in this work with the intention of converting improved forecasts into more intelligent, timely decisions for resilience and adaptation. A Long Short-Term Memory (LSTM) network, an architecture ideal for identifying minute, distant patterns in climate data, lies at the heart of the method. Before the model was taught to identify intricate correlations between climatic variables throughout time, historical climate records were thoroughly cleaned and normalized. In order to evaluate real-world reliability, performance was then compared against ARIMA (AutoRegressive Integrated Moving Average) and two LSTM variants: Bidirectional LSTM and Stacked LSTM. It was then stress-tested against harsh weather conditions and unobserved data. The outcomes were evident: LSTM-based models regularly performed far better than ARIMA, with MAE, MSE, and RMSE of 0.0217, 0.0008, and 0.0285 as opposed to ARIMA’s 1.2809, 1.5290, and 1.2365. These results support deep learning’s significant potential as a scalable and dependable method for regional climate forecasting, giving it an advantage in the valley’s climate change adaptation policy-making process.

A hybrid TFN-LSTM model for groundwater level forecasting in Rhode Island, USA.

Accurate groundwater level forecasting is essential for effective groundwater resource management; however, it remains challenging due to the complex and nonlinear responses of aquifer systems to climatic forcing. Conceptual models, such as transfer function noise (TFN) models, effectively represent long-term trends and seasonal groundwater dynamics; however, they often fail to capture short-term variability, resulting in temporally correlated simulation residuals. Data-driven approaches can capture nonlinear dynamics but often lack clear physical meaning. This research proposes a hybrid model for forecasting groundwater levels by combining a TFN model with a Long Short-Term Memory (LSTM) network through a residual learning approach. Initially, the TFN model simulates baseline groundwater level responses based on precipitation and potential evapotranspiration. Then, the LSTM model is trained to identify temporal patterns in the TFN residuals and predict corrections. These corrections are added to the baseline simulation to produce the final forecasts. The model is tested using daily groundwater level data from three monitoring wells in Rhode Island, USA. The findings indicate that the hybrid model consistently performs better than both the standalone TFN and LSTM models, especially at locations with complex groundwater behavior. Additionally, the hybrid model exhibits greater temporal stability, with its performance declining more gradually over extended forecast periods. Shapley additive explanations reveal that past TFN residuals play a major role in the correction process, emphasizing the significance of residual persistence in improving prediction accuracy.

Typhoon multi-task prediction based on displacement residual learning

Aiming at the limitations that deep learning methods rarely integrate large-scale environmental field information in typhoon track and intensity prediction, and existing models have overfitting risks in absolute coordinate prediction, this paper proposes a residual prediction method for typhoon multi-task prediction. First, based on three architectures including Long Short-Term Memory (LSTM), Transformer and Multilayer Perceptron (MLP), a residual prediction strategy is introduced to convert absolute coordinate prediction into displacement increment prediction. Second, a multi-source dataset is constructed using the International Best Track Archive for Climate Stewardship (IBTrACS) and ERA5 reanalysis data, and 6-hour short-term prediction experiments are designed. With Mean Distance Error (MDE) and classification accuracy as the core evaluation indicators, performance comparison with baseline models is conducted. Experimental results show that the proposed method performs excellently in all models. Among them, the MLP Baseline achieves the optimal track prediction performance (MDE=46.47 km) and the optimal intensity classification performance (accuracy=91.62%); the LSTM+ERA5 fusion model reduces the track MDE by 4.65% compared with the LSTM Baseline, verifying the effectiveness of environmental field feature fusion. The proposed method effectively solves the problem of insufficient utilization of environmental field features in short-term typhoon prediction and provides efficient and accurate deep learning technical support for typhoon disaster early warning.

Explainable Deep Learning Models for Long-Term Temperature and Rainfall Forecasting

Climate impact assessment, agricultural planning, and disaster preparedness requires accurate long-term prediction of temperature and rainfall. Conventional statistical models are frequently not able to represent nonlinear and long-term relationships in climate data, and deep learning models, however powerful in prediction, cannot be interpreted. The given paper suggests an interpretable deep learning architecture, which uses Long Short-Term Memory (LSTM) networks and SHAP-based explainability to forecast multivariate climate. Sliding window sequences are used to process historical data (temperature, rainfall, humidity and atmospheric pressure) to obtain seasonal and long-term trends. The proposed model predicts both temperature and rainfall simultaneously, and it has a better performance than the ARIMA, LSTM, GRU and CNN-LSTM models in the terms of RMSE, MAE and Accuracy. Moreover, the explainability module offers insights into feature and time importance, which increases the level of transparency and trust in the prediction. Empirical evidence shows that the described method is effective in balancing predictive accuracy of 92% and understandability, which can be used in climate decision support systems.