Constraint-Aware Wind Power Forecasting with an Optimized Hybrid Machine Learning Model

Constraint-Aware Wind Power Forecasting with an Optimized Hybrid Machine Learning Model

Accurate prediction of wind power generation (WPG) under real-world scenarios is imperative for achieving optimal costing, ensuring reliable operation, and fortifying the security of power systems. While existing research has proposed numerous single, ensemble and hybrid AI model to enhance prediction accuracy, these architecture often overlook operational constraints. In response, this paper introduces a novel constraint aware forecasting framework formed by a convolutional neural network (CNN) integrated with a double layer of gated recurrent unit (GRU) and fully connected layers. A customized loss function enforces ramping and capacity limits through penalty coefficients, which are optimized using a genetic-adaptive-moment-optimizer (GAMO). On top of that, the performance of the proposed scheme was assessed under diverse ramping threshold settings, ranging from the most stringent worst-case scenarios to relaxed operational conditions. Extensive evaluations on WPG dataset reveal that under the most stringent 10% ramping threshold, the proposed model achieves a MAPE of 3.65%, surpassing Bi-LSTM and CNN models by 7.89% in forecasting accuracy. Additionally, the proposed optimization techniques were bench-marked against the Bayesian optimization process (BOP) with a tree prazen sampler and particle swarm optimization (PSO). The proposed GAMO outperformed these methods in computational efficiency, reducing computation time by 74.20% compared to BOP and 90.91% compared to PSO. Furthermore, the results indicate that the GAMO optimization framework facilitated smoother convergence in the model learning process.