Wind Power Prediction and Uncertainty Modeling for Power System Operation

Abstract

Wind power (WP) is one of the most popular and fastest-growing renewable energy sources in the world. However, the increasing penetration of WP, due to its stochastic nature, introduces an unprecedented level of uncertainty into power systems, potentially jeopardizing grid reliability and economical operation. Effective wind power prediction (WPP) models should be developed to address this challenge, and new methodologies should be designed to handle WP uncertainty in power system scheduling and planning problems. This thesis contributes to WPP and uncertainty modeling by proposing two new probabilistic frameworks for short-term WPP, a new framework for medium- to long-term WPP, and introducing a new approach to incorporate WP uncertainty into generation scheduling. The WP prediction interval (WPPI) model is one of the most common and visualized probabilistic approach to represent WP uncertainty. The proposed short-term WPP frameworks are both developed to generate high-quality prediction intervals (PIs) for the future WP. The first WPPI framework is designed based on the predictive density estimation (DE) of WP to improve prediction performance while simultaneously reducing problem complexity. In contrast to the majority of previous WPPI models, the proposed framework does not optimize the WPPI via a high-dimensional optimization problem with model parameters as control variables. Instead, it optimizes the WPPI by adjusting a single control variable, bandwidth (BW) of the DE. This framework also proposes a new two-stage hybrid deterministic/probabilistic prediction structure that significantly enhances the accuracy and flexibility of the conventional hybrid approach. The second proposed WPPI framework is developed based on an ensemble technique tuned via linear programming. This framework integrates a variety of probabilistic prediction methods with different characteristics, thereby increasing the model's robustness and performance. The proposed ensemble WPPI (EWPPI) model is tuned using a novel two-layer optimization approach. The outer layer of the optimization maximizes the final WPPI quality by adjusting a hyper-parameter used in the inner layer. The inner layer is responsible for optimally weighting each WPPI model within the ensemble prediction engine (EPE) using a proposed linear objective function that maximizes WPPI quality while also taking its symmetricity into account.

For medium- to long-term WPP, this thesis proposes a two-stage prediction framework based on the indirect WPP strategy; the first stage predicts wind speed and direction, and the second stage estimates the available WP in the future time steps according to the power curve and the predicted wind speed and direction values. The wind speed and direction prediction model is based on numerical weather prediction (NWP) data and leverages deep learning to produce high-quality long-term forecasts. Following the decomposition of the historical data, this model employs stacked autoencoders for dimension reduction and hierarchical long short-term memory (LSTM) networks for prediction. The second stage estimates WP according to the deterministic and probabilistic wind power curves. A machine learning-based method is suggested to model the power curve deterministically, and a new straightforward approach is designed for probabilistic power curve modeling. This thesis also introduces a new approach for WP uncertainty modeling, incorporated into power system generation scheduling. The proposed method characterizes the WP uncertainty using a decomposition-based approach and develops a new formulation for generation scheduling with a probabilistic cost function for WP generation. The proposed probabilistic WP cost function provides an estimation of the cost required to compensate for the WP scheduling error. To provide a realistic estimation, a machine learning-based model is developed, which estimates the compensation cost based on the expected values of the overestimation and underestimation with respect to the WP. The overestimation and underestimation expected values are determined based on the historical performance of the WPP model. Also, to estimate the compensation cost in the proposed approach, for the first time, the impact of the load demand, as an effective parameter in the compensation cost, is taken into account, and it is considered as an input to the machine learning-based model to obtain a more accurate cost estimation.