|
|
 |
| |
|
|
|
An Adaptive Error-Controlled LADM–ANN Framework for Nonlinear Fractional Partial Integro-Differential Equations with Weakly Singular Memory Kernels |
|
|
|
PP: 1403-1439 |
|
|
doi:10.18576/jsap/150513
|
|
|
|
Author(s) |
|
|
|
Amin M. Elfeky,
Essa Melfi Aldosari,
Mohamed Sayed Abdellatif,
Mohammad A. Tashtoush,
Hasanain J. Alsaedi,
Adel S. Hussain,
|
|
|
|
Abstract |
|
|
| Fractional partial integro-differential equations (FPIDEs) with weakly singular kernels provide mathematical models for systems exhibiting nonlocality and memory. However, conventional decomposition methods often lack adaptive truncation, problem-specific error control, and reliable long-time continuation. This study develops an adaptive, error-controlled hybrid Laplace-Adomian decomposition and artificial neural network (LADM-ANN) framework combining analytical approximation with physics-guided residual learning. The methodology incorporates model-specific convergence conditions, truncation-error estimates, a dual stopping criterion, and memory-preserving Caputo continuation. Unknown fractional and physical parameters are calibrated through constrained residual minimization, while block-bootstrap inference, sensitivity analysis, and conformal prediction assess uncertainty. Manufactured solution experiments demonstrate substantial numerical improvements, with hybrid RMSE values of 3.396 × 10−4 and 1.402 × 10−5, representing reductions of 83.9% and 89.7%, respectively, relative to adaptive LADM. Real-data validation employs 999,061 filtered NGSIM US-101 trajectory observations aggregated onto a time–space grid. At the 5-s forecasting horizon, the hybrid model achieves an RMSE of 6.149 vehicles/km/lane, outperforming adaptive LADM (6.630) and persistence (6.753). However, at 25 s, its RMSE increases to 22.156, exceeding persistence (20.924). Although the hybrid remains more accurate than persistence at 50–100 s, histogram gradient boosting performs better at these longer horizons. Furthermore, wide bootstrap intervals indicate weak practical parameter identifiability, while conformal under coverage reveals limitations under nonstationary traffic conditions. These findings establish the proposed framework as an interpretable and error-controlled approach with demonstrated short-horizon predictive benefits, while emphasizing that adaptive numerical accuracy does not guarantee superior long-horizon forecasting performance. |
|
|
|
|
 |
|
|