Short-Term Power Load Forecasting Model Based on an Adaptive Multi-Strategy Gold Rush Optimizer for Optimizing an Dilated BiGRU

Short-term power load forecasting plays a crucial role in the operation scheduling, energy management, and security assessment of smart grids, as its accuracy directly affects the economic efficiency and reliability of power systems. However, power load series are characterized by strong nonlinearity, non-stationarity, and multi-scale temporal dependencies, which make it difficult for traditional forecasting models to achieve both high accuracy and strong generalization under complex load scenarios. In recent years, deep learning models have demonstrated promising performance in load forecasting; nevertheless, their effectiveness is highly dependent on hyperparameter configurations. Manual hyperparameter tuning is not only time-consuming but also prone to premature convergence to suboptimal solutions. To overcome these challenges, this study develops an AMS-GRO-based short-term power load forecasting framework. By improving the original Gold Rush Optimizer (GRO), three adaptive strategies are integrated into the proposed optimizer, namely the current-to-pbest/1 mutation strategy, adaptive elite-guided search strategy, and success-rate-based adaptive strategy selection mechanism. These improvements jointly enhance the global exploration capability, local exploitation ability, and strategy adaptability of the optimizer. Subsequently, the proposed AMS-GRO is employed to perform global hyperparameter optimization for a Dilated BiGRU-Attention model, forming the AMS-GRO–Dilated BiGRU forecasting framework. Key hyperparameters, including the learning rate, number of hidden units, attention dimension, and regularization coefficient, are adaptively optimized. Extensive numerical experiments were conducted on the 30-dimensional CEC2017 suite and the 10- and 20-dimensional CEC2022 suites using 30 independent runs. According to the Friedman test, AMS-GRO achieved mean ranks of 1.37, 1.33, and 1.25, respectively, ranking first in all three experimental settings. In the short-term power load forecasting experiment, AMS-GRO–Dilated BiGRU achieved an MAE of 21.09, MAPE of 0.0177, MSE of 791.10, and R2 of 0.9746. Compared with the unoptimized Dilated BiGRU model, these results correspond to reductions of 13.0%, 13.7%, and 21.5% in MAE, MAPE, and MSE, respectively, together with a 0.70-percentage-point improvement in R2. From a symmetry perspective, the proposed framework balances global exploration and local exploitation through adaptive strategy selection, while the Dilated BiGRU exploits bidirectional temporal symmetry to capture multi-scale load patterns. These results demonstrate that the proposed framework provides competitive optimization performance and improves forecasting accuracy on the investigated load dataset.