In July 2021, massive floods across Germany, Belgium and the Netherlands resulted in more than 220 deaths and more than EUR43 billion of losses. Limited capabilities of conventional hydrological and numerical weather prediction (NWP) models to provide accurate short-term flood prediction in highly non-linear and dynamic climatic conditions were revealed by the disaster. This paper presents case analysis of selected as an AI approach to flood forecasting Long Short-Term Memory (LSTM) neural networks, applied to the Ahr Valley basin in Germany which was one of the most severely affected areas. Using the historical hydrological data from 2009 to 2020, stored in the Copernicus Emergency Management Service (CEMS) and the European Centre for Medium-Range Weather Forecasts (ERA5) reanalysis datasets, the model was trained to come up with river discharge levels and flood occurrence probabilities. The LSTM model architecture consisted of four hidden layers each containing a total of 128 neurons, an adaptive learning rate and a mean squared error (MSE) loss function with the Adam optimizer. Results of evaluation against traditional NWP-based forecasts indicated that the LSTM was able to reach prediction accuracy of 92%, root mean square error (RMSE) of 0.13 m and approximately 36 hours increase in the lead time of forecasts. The results show how deep learning approaches can greatly improve flood forecasting accuracy, minimize false alarms and give important early warning for mitigating disasters. This study highlights the possibility of using AI-based hydrology models in the disaster management system of a country to enhance climate resilience and promote Sustainable Development Goals (SDG 11 and 13).