Hybrid intelligent optimisation for onshore wind farm forecasting

Abstract

Accurate wind power forecasting is crucial for the dependable functioning and strategising of contemporary power systems, especially as the global integration of renewable energy escalates. This study introduces an innovative hybrid intelligent forecasting model that amalgamates Long Short-Term Memory (LSTM) neural networks with Complementary Ensemble Empirical Mode Decomposition (CEEMD) and a hybrid optimisation strategy that incorporates Ant Colony Optimisation (ACO), Genetic Algorithm (GA), and Particle Swarm Optimisation (PSO). The model was developed and evaluated utilising empirical data from a 138 MW wind farm consisting of 46 turbines, based on operational data from 2019. The proposed CEEMD-LSTM-ACO-GA-PSO model adeptly tackles the nonlinearity and intermittency of wind speed data through the decomposition of intricate signals, the enhancement of temporal learning, and the optimisation of model hyperparameters. The evaluation results indicated a substantial enhancement in forecasting precision relative to baseline models. The hybrid model attained a Root Mean Square Error (RMSE) of 0.142 and a Mean Absolute Percentage Error (MAPE) of 3.8% for 24-h forecasts, representing an enhancement of more than 35% compared to traditional LSTM models. It also exhibited strong performance over extended forecasting periods of up to 168 h. This study validates the effectiveness of a hybrid intelligent model in improving wind power forecasting while emphasising the limitations associated with computational complexity, sensitivity, feature importance and generalisation. Future research should incorporate uncertainty quantification, simplify models for real-time deployment, and adopt transformer-based architectures. The results endorse the application of intelligent optimisation in enhancing the reliability and sustainability of energy system operations.