EARLY DETECTION OF RAINFALL ANOMALIES USING LSTM AND ISOLATION FOREST

Abstract

The uncertainty of daily rainfall patterns in Cilacap Regency with extreme variations makes it difficult to detect hydrological anomalies early using traditional methods. This study aims to obtain the most optimal LSTM parameters for rainfall prediction models, evaluate model performance using the Mean Squared Error (MSE) also Root Mean Squared Error (RMSE) metric, and predict rainfall anomalies for the next year using Isolation Forest. Daily BMKG data from January 2015 to December 2024 were processed through preprocessing stages, including missing data handling and time sequence creation. The Long Short Term Memory model was trained using regularization techniques to avoid overfitting, and the prediction results were analyzed using Isolation Forest to identify anomalies. The experiment showed that the best combination of hyperparameters was LSTM with 50 units and a tanh activation function, dropout 0.3, followed by a first dense layer of 50 units with ReLU activation and a single output layer. This configuration resulted in a validation MSE of 23.5247161 and RMSE 4.8502, this result identified five cases of anomalies that were validated for suitability based on rainfall data from BPBD. These results demonstrate the model's ability to reconstruct rainfall patterns and detect potential anomalies, so the system has potential to support early warning efforts for disaster mitigation and water resource management in Cilacap.