作者
ABDULLAHI KIRE MUAZU,Siti Maherah Hussin,Md Pauzi Abdullah,Dalila Mat Said,Babangida Modu
摘要
Accurate short-term forecasting of photovoltaic, wind, and electrical load profiles is essential for the stable and cost-efficient operation of renewable-rich power systems. However, the nonlinear and non-stationary characteristics of renewable generation, changing operating conditions, and the lack of unified multi-energy forecasting frameworks remain major challenges. To address these limitations, this study proposes an integrated forecasting-to-operation framework combining complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), multi-task artificial neural network (MTL-ANN) learning, and an online recursive least squares (RLS) adaptation mechanism coupled with a model predictive control-based energy management system (MPC-EMS). By exploiting shared temporal relationships among photovoltaic, wind, and load profiles, the proposed multi-task architecture simultaneously forecasts multiple energy resources while reducing redundancy associated with independently developed forecasting models. The RLS adaptation layer provides lightweight correction of forecasting deviations without requiring complete model retraining, improving robustness under changing operating conditions. Simulation results demonstrate that the proposed framework achieves high forecasting accuracy, with MAPE values of 6.64%, 4.09%, and 0.62% and R² values of 0.998, 0.955, and 0.997 for photovoltaic, wind, and load forecasting, respectively. Compared with the static CEEMDAN-MTL-ANN model, the online RLS adaptation mechanism reduces the MAPE from 7.91% to 6.64% for photovoltaic forecasting, a 16.1% relative reduction that is statistically significant according to the Wilcoxon signed-rank test (p = 2.98 × 10⁻⁷), and by 1.7% and 1.6% for wind and load prediction, respectively. Under synthetic concept drift scenarios representing a 20% PV generation drop and a 15% load demand surge, the RLS adaptation mechanism reduces the post-drift MAPE from 25.66% to 16.77% for PV generation and from 13.03% to 5.63% for load demand, corresponding to relative error reductions of 34.65% and 56.79%, respectively. Furthermore, the integrated MPC-EMS achieves a 3.02% reduction in 24-hour operating cost compared with baseline operation without battery storage, while maintaining 0.00% renewable-energy curtailment and battery state-of-charge within the operational range of 20-90%. Empirical latency evaluation of the LP-based MPC optimization over 100 trials shows an average solving time of 23.30 ms and a worst-case solving time of 1020.28 ms, approximately 0.11% of the 15-minute scheduling interval, confirming the computational suitability of the proposed framework for real-time renewable-energy management applications. The proposed framework provides a computationally efficient and adaptive solution for renewable-rich power system forecasting and operational optimization.