Advancing Stochastic Wind Speed Forecasting Methods with Novel Hybrid Deep Learning Techniques

Advancing Stochastic Wind Speed Forecasting Methods with Novel Hybrid Deep Learning Techniques

Excessive use of fossil fuels intensifies climate change and depletes nonrenewable resources. Transitioning to renewable energy (RE) presents a sustainable solution to mitigate the climate change impacts and reduce the strain on fossil fuel reserves.AmongREsources, wind energy stands out as a promising option, particularly for small islands like Fiji, given its eco-friendly nature, abundant supply, and costeffectiveness. However, wind speed (WS) variability poses a significant challenge for stable wind energy generation. To resolve this, the study develops hybrid deep learning (DL) methods for accurate, reliable, and trustworthy near real-time and short-term WS forecasts. The first objective (publication 1) uses ground-based mereological data of neighbouring stations to develop a bidirectional long short-term memory (BiLSTM) model for forecasting near real-time (10-minute) WS of target stations. To avoid irrelevant predictors from adding a bias to BiLSTM, a multi-stage feature selection (FS) is used. BiLSTM hyperparameters (HPs) are optimized using an efficient Bayesian Optimization (BO). In the second objective (publication 2), a convolutional neural network (CNN) is fused with BiLSTM (CBiLSTM) to predict the short-term (1-hour) WS using diverse ground and satellite-based predictors. Irrelevant inputs are removed using a grey wolf optimizer with a two-phase mutation (TMGWO). Additional improvement to CBiLSTM is achieved using Bayesian Optimization and HyperBand (BOHB). For operational purposes, model interpretability is crucial and is established using eXplainable Artificial Intelligence (xAI) tools: Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). The third objective (publication 3) employs a multivariate empirical mode decomposition (MEMD) tool for simultaneous decomposition ofmultivariate ground-based inputs. An opposition-based whale optimization algorithm(OBWOA) is used for optimal input selection. The selected inputs are used to develop a gated additive tree ensemble (GATE)model formulti-stepahead WS predictions. Overall, the outcomes of this doctoral research are expected to lead to an on-site decision-support framework for wind energy industries. Reliable WS forecasts through the proposed models would help the grid operators optimize wind power grid connections and ensure the safety and stability of power grid operation to mitigate the risk of unwarranted power outages and brownouts. The availability of robust predictive technologies would encourage investors and energy companies to fundmore wind power projects, which will help increase the share of RE in the overall energymix.

File reproduced in accordance with the copyright policy of the publisher/author/creator.