A Multi-Source AI Framework for Intelligent Construction Schedule and Cost Forecasting

In construction analytics, anomaly detection and performance forecasting are ordinarily developed as isolated models. When hybrid models do combine them, the underlying coupling mechanism is seldom detailed sufficiently to allow exact independent replication. Rather than claiming novelty for feature fusion itself, this paper introduces a fully specified, reproducible coupling protocol designed to systematically integrate an anomaly-monitoring layer into a predictive forecasting engine.

Protocol Architecture

Label-Free Online Monitoring: Operates an online interface that exposes the reconstruction error of a CNN-LSTM autoencoder—trained strictly on normal operational data—as a continuous severity score ($\varepsilon_t$) and a binary flag, eliminating any reliance on labeled anomaly ground truth.

Matched Dimensional Embedding: Concatenates these state variables directly into a Transformer predictor’s input embedding layer at matched dimensionality.

Mandatory Paired Ablation: Mandates comparative testing against an identically configured, unconditioned baseline predictor to prove that performance gains originate specifically from the conditioning link rather than raw model capacity.

Latency-Based Evaluation: Measures advance-warning lead time (reported in days relative to equivalent earned-value management signals) directly alongside conventional predictive accuracy metrics.

Empirical Validation

The protocol was validated across three contrasting mega-construction projects in Iraq—an institutional high-rise, a mega-scale residential development, and a marine infrastructure build—representing a combined contract value of USD 15.57 billion. The experimental dataset comprises weekly observations over 24 project-months across 24 input channels. To rigorously eliminate inter-project data leakage, the evaluation utilized leave-one-project-out cross-validation across three folds, training on two projects while evaluating on the held-out third. Crucially, all normalization statistics, anomaly thresholds, and temporal window lengths were fitted strictly within each training fold.

Scope and Limitations

Although the underlying protocol is theoretically applicable wherever a monitoring layer yields scalar or categorical states at decision time, validation across three mega-projects within a single geographic context does not establish universal transferability. Testing the framework on smaller, shorter-duration projects and evaluating its utility across non-construction monitoring domains remain essential targets for future empirical research.