Development of Maintenance Cost Estimation Model using Artificial Neural Network

Abstract

Conventional estimation of building maintenance costs is often inaccurate due to
its reactive and non-data-driven approach, leading to imprecise budgeting
processes. This research aims to examine the potential of using machine learning
with the Artificial Neural Network (ANN) method as a more reliable estimation
tool, and to identify the significant work components influencing the total cost.
Using a case study at the CIBE Building ITB with historical cost data from 2019 to
2024, this study employs the Cost Significant Items (CSI) method to select the most
financially dominant input features, which are further validated using statistical
tests. The prediction model is developed using a feedforward network ANN
architecture with an 8-6-1 configuration, leveraging historical data with a two-year
time lag. To ensure optimal and replicable results, optimization is performed
through the selection of the best random seed. Statistical analysis results indicate
that the Lift Work component is the factor with the most statistically significant
influence on the total maintenance cost (p < 0.05). Meanwhile, the ANN modeling
results demonstrate very high performance, with an accuracy rate on the test data
achieving a Mean Absolute Percentage Error (MAPE) of 0.029%. These findings
strongly suggest that the application of ANN holds significant potential to transform
reactive maintenance budgeting practices into more proactive, accurate, and data
driven approaches.