Energy-efficient forecasting is critical for sustainable grid operation in resource-constrained environments. While artificial neural networks (ANNs) achieve strong predictive performance in short-term wind speed forecasting, their computational demands conflict with Green AI objectives. This paper investigates whether spiking neural networks (SNNs) can attain comparable accuracy at substantially lower inference energy. Using hourly meteorological data from a real-world weather station in Palestine, we systematically benchmark LIF, current-based (CUBA), and sigma-delta (ΣΔ) neurons under rate, time-to-first spike (TTFS), ΣΔ, and direct encodings, trained with supervised surrogate-gradient methods and unsupervised DECOLLE, against fully connected and LSTM baselines under matched architectures and identical protocols. Across five independent runs, supervised SNN configurations define the accuracy–energy frontier. A CUBA model with TTFS encoding at 64 simulation steps achieves MAE ≈ 0.22 and RMSE ≈ 0.30 with statistically indistinguishable performance from LSTM (p > 0.05), while requiring 2,863 pJ per inference—approximately 91% lower energy. ΣΔ neurons with ΣΔ encoding provide an energy-prioritized alternative with minimal accuracy degradation. These results demonstrate that SNNs enable LSTM-level forecasting accuracy at a fraction of the energy cost, supporting low-power edge deployment in renewable-energy systems.