Fiber To The Home Cluster Development: Discovering a Regression Model to Forecast Fiber To The Home Take-Up Rate Using Fitness, Bias/Error Performance, Stability, and Computational Execution as Model Criteria

• Demonstrates a reliable ANN-ELU-based FTTH Take-Up Rate prediction model. • Applies a multi-criteria evaluation that considers fitness, bias, stability, and speed. • Employs a large-scale national FTTH dataset containing 22 critical internal and external elements • ANN-ELU maintains robust even under large data-volume reduction. • Provides interpretable TUR projections with LIME for practical FTTH planning. Predicting the Fiber-to-the-Home (FTTH) Take-Up Rate (TUR) accurately is crucial for network rollout planning, cluster development prioritization, and effective investment allocation. Although earlier methods have shown encouraging accuracy, they typically lack a thorough assessment of the model's scalability and don't offer comprehensible insights that aid in practical decision-making. Six regression models—RBF, ELM, ELM-RBF, ELM-RBF with Richardson Extrapolation, SVR, and an Artificial Neural Network (ANN) with Exponential Linear Unit (ELU) activation—are developed and evaluated in this study using a sizable national FTTH dataset from Indonesia (2021–2024) that includes 54,822 records and 22 important attributes. A multi-criteria evaluation framework that includes accuracy, bias, variance, computational execution, model stability (Friedman and Nemenyi tests), and Expected Squared Error (ESE) is used to evaluate model performance across two train-test splits (70:30 and 80:20) and three dataset-volume scenarios (full, quarter, and one-eighth). The ANN-ELU architecture (23-32-64-1) exhibits the strongest and most consistent generalization, with the lowest squared bias and an ESE of 688.125, and it remains robust even after significant data volume reduction. Beyond predictive accuracy, the ANN's outputs may be locally interpreted with LIME to explain the contribution of internal and external FTTH factors to individual TUR projections, allowing practitioners to understand not only what the model predicts but also why. This interpretability improves the model's usefulness for FTTH cluster planning. Overall, the work provides a transparent multi-criteria assessment demonstrating that ANN-ELU performs similarly to recent gradient-boosting models while maintaining a shorter execution time and better interpretability. These findings support the use of ANN-based predictive analytics to improve FTTH cluster design, rollout prioritization, and homepass investment strategies.