Dynamic cost prediction of power technical transformation projects based on graph neural networks and multi-scale attention mechanisms

Accurate cost prediction for power technical transformation (PTT) projects is central to capital budgeting and investment decisions, yet conventional methods struggle with the tangled interdependencies among cost drivers and with cost dynamics that unfold at several different time scales. We propose a hybrid framework that couples graph neural networks (GNNs) with multi-scale attention for dynamic PTT cost prediction. The method first builds a heterogeneous association graph in which equipment, material, labour, and policy elements appear as interconnected nodes; multi-layer graph attention convolutions with residual connections then extract topology-aware embeddings. Those embeddings feed a three-branch multi-scale attention module with gated fusion that captures short-, medium-, and long-term dynamics at once. On 1,862 real PTT records spanning 2015-2023, the model reaches a mean absolute percentage error of 6.53% and a coefficient of determination of 0.965, outperforming multiple linear regression, support vector regression, a backpropagation network, an LSTM, a Transformer (Informer) baseline, and a GNN-only variant on every metric. Diebold-Mariano tests and paired bootstrap intervals confirm that these gains are statistically significant. Ablation studies show that graph-based relational encoding and multi-scale temporal modelling each contribute, with the GNN identified as the single most influential component.