<p>Industrial control systems (ICS) are increasingly targeted by sophisticated cyber-physical attacks whose signatures are highly nonlinear and spread across multiple temporally dependent sensor streams. Conventional attack detection models often fail to capture heterogeneous temporal and structural dependencies, which reduces their sensitivity to coordinated and slow-drift attack patterns. To address this limitation, we propose a supervised attack detection approach, the Temporally Enhanced Kolmogorov-Arnold Network (TEKAN), which combines spline-based temporal encoding with hypergraph structural reasoning to jointly model functional, temporal, and higher-order sensor relationships. The proposed framework is tested on two water-sector ICS benchmark datasets, SWaT and WADI, using accuracy, precision, recall, F1-score, root mean squared error (RMSE), and mean absolute error (MAE), with training and validation loss used to assess convergence behavior. TEKAN achieves an F1-score of 98.50% on SWaT and 99.56% on WADI. These results show improvements of 1.06 percentage points in F1-score, 1.03 percentage points in recall, and 1.06 percentage points in precision compared with the baseline models. The contribution of key components is examined through controlled ablation analysis using paired <i>t</i>-tests and bootstrap confidence intervals. The findings indicate that TEKAN achieves consistent and improved performance for cyber-attack detection in water-sector ICS scenarios.</p>