A Dual-Phase Metaheuristic CSO-GWO Optimized XGBoost for Household Load Forecasting

Precise short-term prediction of domestic electricity demand is an important precondition of smart grid work, demand-side control, and decentralized energy planning. However, minute-scale residential load data are complex by nature because of the high level of nonlinearity, sharp temporal dynamics, lack of data, and vulnerability to human behavioral patterns. Such properties make it difficult to apply classic statistical methods and machine learning models in situations where hyper parameters are set in a manual or exhaustive search approach. This paper will provide a two-stage metaheuristic optimization system that combines Cat Swarm Optimization and Grey Wolf Optimization to implement systematic tuning of an XGBoost regression model in order to overcome these obstacles. The proposed framework will be intended at balancing global exploration and local exploitation in the hyper parameter lookup space. It improves the firmness of convergence. It decrease the vulnerability of sub optimal parameter configuration. The UCI Individual Household Electric Power Consumption dataset is used to evaluate the methodology, with a chronological train-test split to consider realistic forecasting conditions and avoid being able to use time like information. In experiments, it is shown that the optimized model has a Root Mean Square Error of 0.032951, a Mean Absolute Error of 0.020163, a Mean Absolute Percentage Error of 3.848799 percent, and a coefficient of determination of 0.998628 on unseen data. Further residual diagnostics and convergence testing indicate the low bias, controlled variance, and consistent generalization to be used in different load intensities. The importance analysis of the features further shows that the effect of global intensity is the strongest on the household power demand. Results show that dual-phase metaheuristic optimization is a great way to improve the robustness and predictive reliability of ensemble learning models, which is a computationally feasible and accurate solution to short-term household load forecasting in smart grid settings