Optimized hybrid ensemble learning approaches applied to very short-term load forecasting