Forecasting Models for Integration of Large-Scale Renewable Energy Generation to Electric Power Systems

Abstract

Amid growing concerns about climate change and non-renewable energy sources deple¬tion, vari¬able renewable energy sources (VRESs) are considered as a feasible substitute for conventional environment-polluting fossil fuel-based power plants. Furthermore, the transition towards clean power systems requires additional transmission capacity. Dynamic thermal line rating (DTLR) is being considered as a potential solution to enhance the current transmission line capacity and omit/postpone transmission system expansion planning, while DTLR is highly dependent on weather variations. With increasing the accommodation of VRESs and application of DTLR, fluctuations and variations thereof impose severe and unprecedented challenges on power systems operation. Therefore, short-term forecasting of large-scale VERSs and DTLR play a crucial role in the electric power system op¬eration problems. To this end, this thesis devotes on developing forecasting models for two large-scale VRESs types (i.e., wind and tidal) and DTLR. Deterministic prediction can be employed for a variety of power system operation problems solved by deterministic optimization. Also, the outcomes of deterministic prediction can be employed for conditional probabilistic prediction, which can be used for modeling uncertainty, used in power system operation problems with robust optimization, chance-constrained optimization, etc. By virtue of the importance of deterministic prediction, deterministic prediction models are developed. Prevalently, time-frequency decomposition approaches are adapted to decompose the wind power time series (TS) into several less non-stationary and non-linear components, which can be predicted more precisely. However, in addition to non-stationarity and nonlinearity, wind power TS demonstrates chaotic characteristics, which reduces the predictability of the wind power TS. In this regard, a wind power generation prediction model based on considering the chaosity of the wind power generation TS is addressed. The model consists of a novel TS decomposition approach, named multi-scale singular spectrum analysis (MSSSA), and least squares support vector machines (LSSVMs). Furthermore, deterministic tidal TS prediction model is developed. In the proposed prediction model, a variant of empirical mode decomposition (EMD), which alleviates the issues associated with EMD. To further improve the prediction accuracy, the impact of different components of wind power TS with different frequencies (scales) in the spatiotemporal modeling of the wind farm is assessed. Consequently, a multiscale spatiotemporal wind power prediction is developed, using information theory-based feature selection, wavelet decomposition, and LSSVM. Power system operation problems with robust optimization and interval optimization require prediction intervals (PIs) to model the uncertainty of renewables. The advanced PI models are mainly based on non-differentiable and non-convex cost functions, which make the use of heuristic optimization for tuning a large number of unknown parameters of the prediction models inevitable. However, heuristic optimization suffers from several issues (e.g., being trapped in local optima, irreproducibility, etc.). To this end, a new wind power PI (WPPI) model, based on a bi-level optimization structure, is put forward. In the proposed WPPI, the main unknown parameters of the prediction model are globally tuned based on optimizing a convex and differentiable cost function. In line with solving the non-differentiability and non-convexity of PI formulation, an asymmetrically adaptive quantile regression (AAQR) which benefits from a linear formulation is proposed for tidal uncertainty modeling. In the prevalent QR-based PI models, for a specified reliability level, the probabilities of the quantiles are selected symmetrically with respect the median probability. However, it is found that asymmetrical and adaptive selection of quantiles with respect to median can provide more efficient PIs. To make the formulation of AAQR linear, extreme learning machine (ELM) is adapted as the prediction engine. Prevalently, the parameters of activation functions in ELM are selected randomly; while different sets of random values might result in dissimilar prediction accuracy. To this end, a heuristic optimization is devised to tune the parameters of the activation functions. Also, to enhance the accuracy of probabilistic DTLR, consideration of latent variables in DTLR prediction is assessed. It is observed that convective cooling rate can provide informative features for DTLR prediction. Also, to address the high dimensional feature space in DTLR, a DTR prediction based on deep learning and consideration of latent variables is put forward. Numerical results of this thesis are provided based on realistic data. The simulations confirm the superiority of the proposed models in comparison to traditional benchmark models, as well as the state-of-the-art models.