A Modular Tide LevePrediction Using Combination Ofharmonic-Analysis and Nonlinear Autoregressive Exogenous (Narx) Methodology in Semarang Indonesia

Abstract

Semarang is highly prone to tidal floods year-round, making tidal prediction using methods like Harmonic Analysis with Least Squares (HA-LS) crucial for disaster mitigation. However, this method is not effective enough because it has relatively high Root Mean Squared Error (RMSE) value of up to 11.43 cm with coefficient of determination (R²) of around 0.727. The Nonlinear Autoregressive Exogenous (NARX) neural network model is then proposed to improve the accuracy of tide predictions. In general, the features used as input are divided into two types, namely atmospheric/weather data (temperature, pressure, direction, and wind speed) and estimated tide data from the HA-LS method. This research aims to analyze whether this type of neural network is suitable to be applied in Indonesia since the ocean and atmosphere condition might vary. Tidal observation data from the Tanjung Mas Maritime Meteorological Station is used as target data/actual data. Three scenarios are used by varying input types to find out which type of input produces the best performance of model prediction. Moreover, before feeding input data into the NARX neural network, all atmospheric data used as input are standardized using Z-score normalization, often called Min-Max scaling which can avoid the effect of outlier in the dataset. Based on these three scenarios, the use of combined atmospheric/weather data and tide estimates from HA-LS calculations as input to the NARX model produces the best predictive performance with the smallest RMSE value among all scenarios, approximately 5 cm, and the highest coefficient of determination (R²) at about 0.974. These results indicate that NARX model can predict tides with high accuracy in Semarang.