Bio-Inspired Metaheuristic Optimization of a DWT-BiLSTM Architecture for Wind Speed Forecasting: A Statistical...

Population-based bio-inspired metaheuristics are the dominant tools for tuning hybrid decomposition-deep-learning forecasters, yet their relative behavior on a common problem is rarely assessed with a leakage-free, physically meaningful protocol. We benchmark eight metaheuristics on the joint nine-dimensional hyperparameter optimization of a discrete-wavelet-transform bidirectional-LSTM (DWT-BiLSTM) architecture for short-term wind speed forecasting, using 409,152 hourly observations from eight meteorological stations. The set comprises six nature-inspired methods (Artificial Bee Colony, ABC; genetic algorithm, GA; Particle Swarm Optimization, PSO; Grey Wolf Optimizer, GWO; Hippopotamus Optimization, HO; and the Raindrop Optimizer) together with two recent metaphor-free or social variants (the Farthest-better Nearest-worse Optimizer, FNO; and the Tuckman Optimization Algorithm, TOA). A multi-stage protocol covers 30 independent runs per algorithm, a joint-versus-sequential comparison, a genuine rolling-origin out-of-sample evaluation, and component ablation. Friedman testing reveals significant differences (χ² = 49.76; p < 10^-8), with the Grey Wolf Optimizer attaining the best mean rank (2.27) and Pareto-dominant run-time; ablation shows the DWT front-end is essential (Cohen's d = 13.09) and bidirectionality negligible at the one-hour horizon (p = 0.674). Critically, evaluating forecasts in reconstructed physical units reveals that the per-component advantage does not persist: at the one-hour horizon the reconstructed forecast does not exceed a naive persistence baseline (skill ≈ -0.5 in m/s versus +0.44 in normalized component space), a discrepancy independent of decomposition leakage that we report transparently. This work thus contributes a rigorous, leakage-controlled bio-inspired benchmark and a cautionary evaluation methodology.