Anaísa Lucena, Ana Martins, Armando J. Pinho et al. · Mar 19, 2026
Autoregressive (AR) models are formulated into a feedforward Neural Network (NN), enabling coefficient estimation through backpropagation while preserving interpretability. The proposed method consistently recovers model coefficients, achieving substantial computational gains of up to 34.2x and comparable estimation accuracy to Conditional Maximum Likelihood (CML) when it converges. The NN-based approach provides reliable estimates even when CML fails to converge.
Why This Matters
This paper's proposed Neural Network-based method for fast and interpretable autoregressive estimation can be particularly useful for power system engineers, as it enables the efficient estimation of model coefficients for short-term dependence scenarios, which are crucial for predicting renewable energy output and optimizing grid operations. The substantial computational gains achieved by this approach can also support real-time forecasting and grid resilience analysis.