PSO-TCWavKAN: A Swarm-Optimized Temporal Convolutional Wavelet Kolmogorov-Arnold Network with Multi-Kernel Trend Decomposition and Asymptotic Linear Lower-Bound Guarantees for Resilient Power Transformer Thermal Dynamics Forecasting

This paper proposes PSO-TCWavKAN, a swarm-optimized temporal convolutional wavelet Kolmogorov-Arnold network for power transformer thermal dynamics forecasting. The architecture integrates continuous dyadic Morlet wavelet filter banks, multi-kernel trend-seasonal decomposition, and a parameterized neural routing gate with asymptotic linear recovery guarantee (Proposition 1). A Comprehensive Learning Particle Swarm Optimizer (CLPSO) conducts global hyperparameter warm-start, followed by AdamW fine-tuning and Stochastic Weight Averaging. Evaluated across 16 canonical ETT benchmark configurations, PSO-TCWavKAN dominates 7/8 horizons on ETTh2, achieves up to 13.6% MSE reduction over DLinear, with 44.8K parameters and 3.4ms inference latency. Code: https://github.com/nphu211206/kan-grid-update-reliability