Tabel 1. Dam sluice gate opening procedure Tabel 2. Details of experiments Figure 2. Flood monitoring system architecture After data was obtained from the sensors then the data was processed. Table 2 shows two scenarios for using the GRU model to predict downstream water levels. Scenario 1 used downstream precipitation data, while Scenario 2 adds water discharge. This study explores two scenarios using the GRU model for flood prediction based on different ratios of training and validation data. Scenario | investigates the influence of rainfall and water levels downstream; each scenario was tested with training to test data ratios of 70:30, 80:20, and 90:10. Meanwhile, Scenario 2 includes water discharge data along with downstream rainfall, also divided in the same ratio. This scenario aims to evaluate the accuracy of model predictions under different training conditions, which is important for optimizing the flood prediction model in East Flood Canal Semarang. Figure 3. Gated recurrent unit Figure 3 shows the architecture of the GRU, which is a type of recurrent neural network (RNN) architecture used to process sequential data. GRU consists of two main gates: the reset gate rt and the update gate z, which control the flow of information between the network units. The reset gate determines how much information from the previous state should be forgotten, while the update gate determines how much new information should be stored in the network state. With this architecture, GRU is able to capture long- term dependencies in sequential data without experiencing the gradient decay problem that often occurs in traditional RNNs. Table 3. Classification report dataset 70:30 Table 4. Classification report dataset 80:20 3.3. Downstream water level prediction analysis dataset 90:10 Figure 5. Comparison of actual vs. predicted water levels using (a) GRU and (b) classification accuracy with an 80:20 dataset split Fl-score is 1.00, based on 138 samples. The “Caution” category shows a precision of 0.97, recall of 1.00, and Fl-score of 0.98 from 28 samples. The “Emergency” category, with 7 samples, has a precision of 0.88, recall of 1.00, and Fl-score of 0.93. The “Error” category, with 33 samples, has a precision of 1.00, recall of 0.97, and Fl-score of 0.98. The overall accuracy of the model is 0.99 from 206 samples. The macro average for precision, recall, and Fl-score is 0.96, 0.99, and 0.97, respectively, while the weighted average for these metrics is 0.99, indicating excellent model performance. Figure 6. Comparison of actual vs. predicted water levels using (a) GRU and (b) classification accuracy with a 90:10 dataset split OE EE Figure 6 shows the comprehensive results of the model performance using dataset of 90:10. It is livided into two sub-figures: Figures 6(a) and 6(b). Figure 6(a) represents a comparison between the actual Jownstream levels and t accurate and agree wel etter, Which deviates from of the true value. The calculated MSE of 0.135 gives t lifference between pred classification model wit! icted and actual values. Figure 6(b) shows an evaluation o h the confusion matrix on dataset 90:10. In this case, the mod lata as “alert” and 17 data as “caution.” However, only 3 data were correctly classi classification errors were recorded in this matrix, indicating high accuracy. he predicted values generated by the model. MAPE with a value of 0.00186 indicates he percentage of the mean deviation between the predicted and actual values. T with the actual values. The RMSE is calculated as 0.36 naximum error between the predicted and actual results. A lower RMSE indicates he predictions are more 7, which represents the the model can perform he difference in squarec f the performance of the el correctly classified 67 fied as “emergency.” Nc Table 5. Classification report dataset 90:10 3.4. Scenario 1 and 2 comparison results 4. CONCLUSION
