Constraint first, shrinking next: A hybrid photovoltaic generation forecasting framework based on ensemble learning and multi-strategy improved optimizer

Abstract

Accurate photovoltaic (PV) power generation forecasting is critical for effective smart grid management. However, PV power is highly dependent on weather variables, resulting in significant volatility and nonlinearity in generation patterns, which makes traditional forecasting methods often struggle to guarantee both accuracy and robustness at the same time. Although Pareto-front-based ensemble learning approaches partially mitigate these issues, the excessive generation of suboptimal solutions increases decision-making challenges for grid operators. To this end, this paper customizes a novel PV ensemble learning forecasting system based on Bi-Directional Guarantee (BIDG) and Pareto Front Shrinking (PFS). Specifically, the original Hippopotamus Optimizer (HO) is enhanced into a Multi-Objective Hippopotamus Optimizer (MOHO) using a crowding-distance-based multi-objective framework. To further enhance the algorithm’s performance, low-discrepancy sequence initialization is adopted to improve the optimizer’s exploration ability, and a perturbation operator is introduced to enhance convergence performance. The BIDG mechanism based on the improved MOHO is used as the weighting strategy for ensemble learning, which imposes dual constraints on the accuracy and robustness of the forecast and penalizes suboptimal solutions to improve the quality of the Pareto front. To further improve the quality of the Pareto front to simplify decision-making, we combine the Shrinking Method to Quantify Competition (SMQC) and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) to shrink the Pareto front. Numerical validation is conducted using seasonal local datasets from a PV plant in Australia’s Northern Territory. The experimental results show that the average absolute percentage error of the hybrid framework on the four datasets are MAE_Spring = 0.3518, MAE_Summer = 0.5689, MAE_Autumn = 0.8585, MAE_Winter = 1.2175, which are significantly lower than other models, achieving higher forecasting accuracy and demonstrating its practical applicability for real-world smart grid operations.