Utilizing LSTM-GRU for IOT-Based Water Level Prediction Using Multi-Variable Rainfall Time Series Data

Abstract

This research describes experiments using LSTM, GRU models, and a combination of both to predict floods in Semarang based on time series data. The results show that the LSTM model is superior in capturing long-term dependencies, while GRU is better in processing short-term patterns. By combining the strengths of both models, this hybrid approach achieves better accuracy and robustness in flood prediction. The LSTM-GRU hybrid model outperforms the individual models, providing a more reliable prediction framework. This performance improvement is due to the complementary strengths of LSTM and GRU in handling various aspects of time series data. These findings emphasize the potential of advanced neural network models in addressing complex environmental challenges, paving the way for more effective flood management strategies in Semarang. The performance graph of the LSTM, GRU, and LSTM-GRU models in various scenarios shows significant differences in the performance of predicting river water levels based on rainfall input. The MAPE, MSE, RMSE, and MAD metrics are presented for training and validation data in six scenarios. Overall, the GRU model and the LSTM-GRU combination provide good performance when using more complete input variables, namely, downstream and upstream rainfall, compared to only using downstream rainfall.