Data-driven modelling, forecasting and uncertainty analysis of disaggregated demands and wind farm power outputs

Abstract

Correct analysis of modern power supply systems requires to evaluate much wider ranges of uncertainties introduced by the implementation of new technologies on both supply and demand sides. On the supply side, these uncertainties are due to the increased contributions of renewable generation sources (e.g., wind and PV), whose stochastic output variations are difficult to predict and control, as well as due to the significant changes in system operating conditions, coming from the implementation of various control and balancing actions, increased automation and switching functionalities, and frequent network reconfiguration. On the demand side, these uncertainties are due to the installation of new types of loads, featuring strong spatio-temporal variations of demands (e.g., EV charging), as well as due to the deployment of different demand-side management schemes. Modern power supply systems are also characterised by much higher availability of measurements and recordings, coming from a number of recently deployed advanced monitoring, data acquisition and control systems, and providing valuable information on system operating and loading conditions, state and status of network components and details on various system events, transients and disturbances. Although the processing of large amounts of measured data brings its own challenges (e.g., data quality, performance, and incorporation of domain knowledge), these data open new opportunities for a more accurate and comprehensive evaluation of the overall system performance, which, however, require new data-driven analytical approaches and modelling tools.

This PhD research is aimed at developing and evaluating novel and improved data-driven methodologies for modelling renewable generation and demand, in general, and for assessing the corresponding uncertainties and forecasting, in particular. The research and methods developed in this thesis use actual field measurements of several onshore and offshore wind farms, as well as measured active and reactive power demands at several low voltage (LV) individual household levels, up to the demands at medium voltage (MV) substation level. The models are specifically built to be implemented for power system analysis and are actually used by a number of researchers and PhD students in Edinburgh and elsewhere (e.g., collaborations with colleagues from Italy and Croatia), which is discussed and illustrated in the thesis through the selected study cases taken from this joint research efforts.

After literature review and discussion of basic concepts and definitions, the first part of the thesis presents data-driven analysis, modelling, uncertainty evaluation and forecasting of (predominantly residential) demands and load profiles at LV and MV levels. The analysis includes both aggregation and disaggregation of measured demands, where the latter is considered in the context of identifying demand-manageable loads (e.g., heating). For that purpose, periodical changes in demands, e.g., half-daily, daily, weekly, seasonal and annual, are represented with Fourier/frequency components and correlated with the corresponding exploratory meteorological variables (e.g., temperature, solar irradiance), allowing to select the combination of components maximising the positive or negative correlations as an additional predictor variable. Convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) are then used to represent dependencies among multiple dimensions and to output the estimated disaggregated time series of specific load types (with Bayesian optimisation applied to select appropriate CNN-BiLSTM hyperparameters). In terms of load forecasting, both tree-based and neural network-based models are analysed and compared for the day-ahead and week-ahead forecasting of demands at MV substation level, which are also correlated with meteorological data. Importantly, the presented load forecasting methodologies allow, for the first time, to forecast both total/aggregate demands and corresponding disaggregated demands of specific load types.

In terms of the supply side analysis, the thesis presents data-driven evaluation, modelling, uncertainty evaluation and forecasting of wind-based electricity generation systems. The available measurements from both the individual wind turbines (WTs) and the whole wind farms (WFs) are used to formulate simple yet accurate operational models of WTs and WFs. First, available measurements are preprocessed, to remove outliers, as otherwise obtained WT/WF models may be biased, or even inaccurate. A novel simulation-based approach that builds on a procedure recommended in a standard is presented for processing all outliers due to applied averaging window (typically 10 minutes) and WT hysteresis effects (around the cut-in and cut-out wind speeds). Afterwards, the importance of distinguishing between WT-level and WF-level analysis is discussed and a new six-parameter power curve model is introduced for accurate modelling of both cut-in and cut-out regions and for taking into account operating regimes of a WF (WTs in normal/curtailed operation, or outage/fault).

The modelling framework in the thesis starts with deterministic models (e.g., CNN-BiLSTM and power curve models) and is then extended to include probabilistic models, building on the Bayesian inference and Copula theory. In that context, the thesis presents a set of innovative data-driven WT and WF probabilistic models, which can accurately model cross-correlations between the WT/WF power output (Pout), wind speed (WS), air density (AD) and wind direction (WD). Vine Copula and Gaussian mixture Copula model (GMCM) are combined, for the first time, to evaluate the uncertainty of Pout values, conditioning on other explanatory variables (which may be either deterministic, or also uncertain). In terms of probabilistic wind energy forecasting, Bayesian CNN-BiLSTM model is used to analyse and efficiently handle high dimensionality of both input meteorological variables (WS, AD and WD) and additional uncertainties due to WF operating regimes. The presented results demonstrate that the developed Vine-GMCM and operational WF model can accurately integrate and effectively correlate all propagated uncertainties, ultimately resulting in much higher confidence levels of the forecasted WF power outputs than in the existing literature.