State-adaptive interval type-2 neuro-Fuzzy inference system for noise-Resilient short-Term load forecasting

• State-adaptive IT2 neuro-fuzzy system for noise-resilient load forecasting. • Dynamic rule weighting enables robust adaptation to concept drift. • Differentiable type-reduction provides O(R log R) real-time scalability. • 83.4 % error reduction vs LSTM across multiple noise scenarios. • Interpretable fuzzy rules support smart grid decision trust. Short-term load forecasting (STLF) is a cornerstone of modern power system management, yet conventional statistical and machine learning models often fail under noisy, incomplete, and non-stationary conditions. This paper proposes a novel State-Adaptive Interval Type-2 Neuro-Fuzzy Inference System (IT2-SANFIS) designed to address these challenges through the integration of adaptive rule evolution, interval type-2 fuzzy sets, and efficient type-reduction. Unlike traditional models, IT2-SANFIS dynamically modifies its rule base in response to evolving consumption patterns and environmental states, while providing interpretable forecasts with quantified uncertainty. The model was evaluated on hourly electricity demand data from Johor, Malaysia, under clean and noisy conditions (Gaussian, speckle, and impulse noise) and benchmarked against some state-of-the-art approaches, including SARIMA, LSTM, and hybrid fuzzy-neural systems. Results demonstrate that IT2-SANFIS achieves superior performance across most benchmarks, with an RMSE of 2.6048 MW on normalized clean data and maintaining robustness with RMSE = 2.8294 MW at 10 % Gaussian noise. Additional validation on AEP and ISONE datasets confirms generalizability, with MAPE improvements of 98.5 % and 95.4 % respectively over the best deep learning models. The model’s state-adaptive rule weighting and differentiable type-reduction provide significant advantages in interpretability and computational efficiency, achieving up to 5.7 × speedup over classical methods. These results underscore the suitability of IT2-SANFIS for deployment in real-world smart grid environments.