Time series forecasting is crucial in predictive maintenance and reliability analysis, particularly for datasets like NASA’s turbofan engine degradation dataset. While traditional algorithms like LSTMs offer high accuracy, their opaque nature limits explainability and adaptability when retrained on new data, often leading to catastrophic forgetting. This research investigates the potential of Kolmogorov–Arnold Networks (KAN) as an alternative to perceptron-based models, leveraging their inherent explainability and retraining flexibility. This study evaluates two KAN-based approaches—vanilla KAN and RMoKE—and compare their performance with multiple MLP-based models. Experimental results demonstrate that RMoKE achieves a mean squared error (MSE) of 1.8078 × 10⁻⁸ and a mean absolute error (MAE) of 1.3324 × 10⁻4. The MAE and MSE outperform LSTM by 67% and 37%, respectively, while using 99.88% fewer parameters than the LSTM model. This study highlights the promise of KANs in advancing both the accuracy and transparency of time series forecasting models.