Gate Spiking Neural Networks with Heterogeneous Synapses and Entropy Regularization for Very Short-Term Wind Power Forecasting

Accurate very short-term wind power forecasting requires both precise point estimates and reliable uncertainty quantification. This paper evaluates an Gate Spiking Neural Network (GS-NN) pipeline for 1-minute-ahead wind power prediction on the Monash wind power benchmark after resampling to 1-minute resolution. The study compares a baseline spiking neural network (SNN) and two composable innovations, HetSyn and Entropy-Max, against tuned eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest baselines. The protocol uses a chronological 60/15/10/15 train/validation/calibration/test split, three random seeds, and split-conformal prediction intervals with target coverage 0.90. The best result is obtained by the HetSyn-based SNN ensemble, which achieves a root mean squared error of 0.0596 and outperforms XGBoost, Random Forest, and Light-GBM. Diebold–Mariano tests indicate significant gains over the tree baselines, while the main remaining limitations are undercoverage relative to the nominal confidence level.