Improving Indonesia's tsunami early warning. Part II: Hybridized deep learning and metaheuristic algorithm for forecasting and optimizing

In the previous part, we discussed deploying tsunami observation networks in South Java. This part advances the work by focusing on developing deep-learning-based tsunami forecasting and novel buoy configuration optimization through hybridized deep learning and metaheuristic algorithms. This aligns with our primary objective, which is to optimize tsunami buoy configurations and forecast waveforms with speed and reliability. From Part I, we extracted tsunami waveforms for all the OBPGs and coastal points at 50-m depth. We preprocessed the time series data and applied the Long- and Short-Term Time Series Network (LSTNet) for our forecasting model. For optimization, we developed the LSTNet-GA, a hybrid model integrating Genetic Algorithms with LSTNet, to determine the optimal buoy configuration based on its performance. The Kneedle algorithm was then adopted to identify the optimal number of sensors. We proposed 26 OBPGs with the specific configuration based on the LSTnet-GA. Our results show the proposed method is robust, reliable, and more computationally efficient than standard methods.