<![CDATA[The Bangka Belitung Islands Province frequently experiences drought, making it difficult to meet the clean water needs of local communities. The clean water crisis has become a significant issue as it repeatedly occurs during the dry season. This problem can be addressed by conducting mapping in the form of water discharge prediction for water sources that have the potential to be developed as raw water sources. One of the most abundant potential raw water sources in the Bangka Belitung Islands Province is the former tin-mining ponds (locally called kolong). One approach to assess the potential of these ponds as raw water sources is by predicting their water discharge. This study aims to predict the water discharge of the Kolong PL reservoir located in Pangkalan Baru District using an Artificial Neural Network (ANN) method with a feedforward backpropagation approach. The input dataset consists of monthly water discharge data from January 2007 to December 2023. The ANN model was developed and compared with empirical approaches to evaluate the effectiveness of neural networks in predicting water discharge. The results show that the Mean Absolute Percentage Error (MAPE) for the training data is very good at 10.10%, while the testing data exhibits weak predictive performance with a MAPE of 62.72%. This disparity indicates that the ANN’s performance is highly dependent on the selection of the network architecture, particularly the optimal number of hidden layers. Determining the appropriate number of hidden layers requires multiple simulations to achieve the configuration that produces high predictive accuracy. To strengthen the validity of the ANN model, a simple comparison was conducted using linear regression and ARIMA. The linear regression model yielded a training MAPE of 34% and a testing MAPE of 70%, while the ARIMA (1,0,1) model produced a testing MAPE of 66%. These results indicate that the ANN does not provide a significant advantage in the testing phase, although its training performance is superior to both linear regression and ARIMA. This study indicates that the ANN method has strong potential for learning water discharge patterns; however, optimization of the network architecture is required to achieve accurate predictions. The implications of this research highlight that developing an optimally configured ANN-based prediction model can serve as a valuable foundation for planning and managing raw water resources in the Bangka Belitung Islands region.]]>
Copyrights © 2026
