GMDH-enhanced temporal convolutional network for short-term wind forecasting and microgrid operation

• Hybrid polynomial network feature construction boosts short-term wind forecast accuracy. • Proposed model cuts 1‑hour wind speed prediction error by 47% versus persistence. • Forecast‑driven model predictive control reduces simulated operating cost by 23.7%. • Simulation achieves near‑zero diesel use in a remote wind‑diesel microgrid. The inherent variability of wind power poses a problem for the stability and cost-effectiveness of remote microgrids. Hence, accurate short-term forecasting is vital for effective energy management. In this paper, a new hybrid approach that utilizes a group method of data handling (GMDH) network for automated nonlinear feature learning and a temporal convolutional network (TCN) for learning long-range temporal dependencies is proposed. The proposed approach is tested on a large set of benchmarks, including persistence, ARIMA, machine learning techniques (random forest, support vector regression), and deep learning techniques (BiLSTM, GRU, Transformer, and TCN). The proposed model outperforms the existing techniques. When tested on a real-world dataset from Bonavista, NL, the proposed approach achieves the minimum errors (MAE: 0.182 m/s, RMSE: 0.272 m/s), which is a 47% reduction in RMSE compared to the persistence model. An ablation study also shows that the proposed approach is significantly better as a result of the inclusion of the GMDH component. When integrated with a model predictive control (MPC) framework for a simulated wind-diesel-battery microgrid, the proposed approach achieves a 23.7% reduction in costs and a 99.9% reduction in diesel consumption. This demonstrates the potential of the model to improve the economics and sustainability of renewable-based microgrids.