Optimization of Earthquake B-Value Prediction in Java Using GRU and Particle Swarm Optimization

Accurate prediction of earthquake parameters is essential for seismic risk assessment and disaster mitigation, particularly in tectonically active regions such as Java Island, Indonesia. This study presents a novel predictive model for estimating the earthquake b-value a fundamental seismological parameter representing the logarithmic relationship between earthquake frequency and magnitude by integrating a Gated Recurrent Unit (GRU) neural network with Particle Swarm Optimization (PSO). The model is trained using earthquake catalog data from 1962 to 2024, sourced from the Indonesian Meteorological, Climatological, and Geophysical Agency (BMKG). The GRU architecture is selected for its effectiveness in modeling temporal dependencies in seismic time series data. PSO is employed to optimize essential hyperparameters, including the number of GRU units, learning rate, and dropout rate. The optimized model achieves notable improvements in predictive performance: Mean Squared Error (MSE) is reduced from 0.00435 to 0.00030, Root Mean Squared Error (RMSE) from 0.0509 to 0.0173, and Mean Absolute Percentage Error (MAPE) from 3.42% to 1.12%. Training time is also reduced from 57 seconds to 33 seconds, indicating greater computational efficiency. The optimal PSO settings include an inertia weight of 0.8, cognitive and social coefficients of 1.0, 40 particles, and 10 iterations. The primary novelty of this study lies in its targeted application of PSO-optimized GRU architecture for b-value prediction in a seismically complex region. These results demonstrate that evolutionary optimization significantly enhances deep learning performance, providing a robust and efficient framework to support earthquake forecasting and risk mitigation efforts in high-risk zones such as Java Island.