DEVELOPMENT OF AI-BASED MICROGRID RENEWABLE GENERATION FORECASTING

Abstract

The growing integration of renewable energy resources into modern microgrids has intensified the need for accurate, scalable, and uncertainty-aware forecasting frameworks capable of operating under dynamic meteorological conditions. Wind and solar generation exhibit strong intermittency, non-stationarity, and complex spatial dependencies, making conventional statistical and deep learning models insufficient for real-time operational decision-making. This thesis develops a comprehensive suite of artificial intelligence–based methodologies that address multi-horizon deterministic forecasting, ensemble-based prediction refinement, uncertainty quantification, and cross-source spatio-temporal modeling tailored for microgrid applications. The research begins with a systematic review of existing renewable forecasting paradigms, emphasizing the evolution from classical physical and statistical models to modern machine learning, deep learning, ensemble learning, and probabilistic approaches. Critical research gaps are identified in multi-scale temporal modeling, hybridization strategies, uncertainty calibration, and the treatment of spatial correlations in hybrid microgrid settings. Motivated by these insights, the first methodological contribution introduces a hybrid deep learning pipeline that integrates PCA- and EMD-based feature engineering with an attention-augmented LSTM model and XGBoost residual refinement. This framework significantly improves short-term wind power forecasting accuracy across multiple real-world datasets from Tamil Nadu, India. Building upon these findings, the thesis proposes the Confined Attention-enabled LSTM (CAELSTM) architecture, which explicitly separates localized short-term dynamics from long-term periodic components via a dual-branch temporal modeling strategy. The confined-attention mechanism restricts the receptive field to the most informative temporal windows, while a parallel periodic extraction module captures diurnal and multi-scale variations. Extensive experiments demonstrate superior accuracy and robustness compared to classical LSTM variants and state-of-the-art hybrid models. To enhance v multivariate forecasting reliability, the research advances an ensemble-based framework combining CEEMDAN decomposition, Stacked GRUs, and a hybrid bagging–boosting strategy. This design effectively mitigates noise, reduces variance, and strengthens generalization under diverse operating regimes, providing a highly stable forecasting alternative for microgrid supervisory control. Recognizing the limitations of deterministic forecasting in operational environments, the thesis develops a Physics-Informed Adaptive Conformal Prediction (PI-ACP) approach for uncertainty quantification. The proposed method integrates regime-aware nonconformity scoring, adaptive calibration via exponentially weighted residuals, and physics-based generation constraints to ensure reliable, distribution-free predictive intervals that remain consistent under temporal drift and wind ramp events. An extended multivariate formulation enables joint multi-horizon interval forecasting for improved operational risk assessment. The final contribution presents a unified spatio-temporal transformer architecture designed for hybrid solar–wind microgrids. The model employs dual-spatial attention to capture geographical and meteorological couplings across distributed nodes, and dual-phase temporal attention to learn both short-term fluctuations and long-term seasonal trends. Coupled with a PI-ACP layer, the framework delivers physically consistent probabilistic forecasts suitable for real-time microgrid deployment. Validation on a six-node hybrid microgrid in the Khavda–Bhuj region demonstrates significant advancements in accuracy, generalization, and probabilistic reliability relative to deterministic, quantile-based, and diffusion-based baselines. Overall, the thesis establishes an integrated forecasting ecosystem combining hybrid deep learning, attention mechanisms, ensemble learning, adaptive probabilistic calibration, and spatio-temporal transformers. The developed methodologies collectively advance renewable generation forecasting for intelligent microgrid operations, supporting improved stability, planning, and decision-making in renewable-rich power systems.