Duehee Lee

Short-Term Wind Power Ensemble Prediction Based on Gaussian Processes and Neural Networks

We propose an ensemble short-term wind power forecasting model that is based on our novel approaches and advanced forecasting techniques in the modern literature. The performance of our model has been verified by forecasting wind power up to 48 hours ahead at seven wind farms for one and a half years. Our model ranked fourth in the Power and Energy Society (PES) wind power forecasting competition. The forecasting model uses 52 Neural Network (NN) sub-models and five Gaussian Process (GP) sub-models in parallel. For 48 hours, the NN sub-models forecast the future wind power based on historical wind power data and forecasted wind information. In parallel, for the first five hours, five GP sub-models are used to forecast wind power using only historical wind power in order to provide accurate wind power forecasts to NN sub-models. These models provide various forecasts for the same hour, so the optimal forecast should be decided from overlapped forecasts by the decision process.

Waveform characterization of animal contact, tree contact, and lightning induced faults

In this paper signal processing tools are used to uncover common and unique characteristics of faults resulting from animal contacts, tree contacts and lightning. For each fault type a large number of voltage and current waveform data sets measured at monitoring stations on distribution systems are analyzed. The characteristics include but are not limited to the presence of impulse-like oscillations, the number of phases involved, the duration of fault event, the phase angle, the time of day, the spectral content in the time-frequency and time-scale domains, the rate of rise of voltage or current, and the arc voltage. An individual characteristic alone is insufficient to provide an estimate of the fault type. However, by combining common and unique characteristics extracted from a fault event, it may be possible to estimate the fault type accurately.

Future Wind Power Scenario Synthesis Through Power Spectral Density Analysis

Scenarios of near future wind power are synthesized by considering the power spectral density (PSD), statistical characteristics, and the future capacity. The PSD of the wind power follows different power laws over different frequency ranges and is approximated by a piecewise function. A scaling exponent of the power law for a particular piece can be approximated by the slope of an affine function fitted to a logarithmic plot of the PSD. Each piece of the function has a different trend as the total capacity increases. Slope trends, the first PSD value, and the last PSD value are trained to forecast the PSD. Then, future wind power scenarios are synthesized from the forecasted PSD. In this process, phase angles are searched using a genetic algorithm while satisfying forecasted statistical characteristics for the given capacity. Our approach is simulated and validated for wind power for seven years in ERCOT and is used to synthesize a future wind power scenario at 10,000 MW capacity. Our approach could also be used to generate wind power scenarios at present capacity for many stochastic optimization problems in power systems.

Stochastic Optimal Control of the Storage System to Limit Ramp Rates of Wind Power Output

The purpose of this paper is to reduce the required amounts of ancillary services, by limiting ramp rates of wind power through control of a storage system. Storage operation policies to limit the ramp rate of net production (wind + storage output) are optimized in a two-stage stochastic linear program with fixed recourse. Operation policies decide the storage operation based on the probability density function of one step ahead wind power, which is forecasted by a Gaussian process. In this optimization problem, it is assumed that a financial penalty is incurred when violating the ramp rate limit (RRL), and that the penalty is linearly proportional to the number of MW per period above or below the RRLs. The L-shaped method is used to reduce the number of constraints caused by various wind power scenarios. Then, the storage specification is determined based on minimizing the sum of the penalty costs and the investment costs of the storage system. The ultimate goal of this paper is to find the relationship between the financial penalties and the RRL.

Analyzing the variability of wind power output through the power spectral density

The variability of wind power output is analyzed using the power spectral density (PSD) that is estimated through the modified covariance method, one of the parametric PSD estimation techniques. This method can estimate the PSD accurately under the assumption that the wind power follows the Autoregressive process. The seasonal trends are removed using the Multiple signal classification to satisfy this assumption before estimating the PSD. Phase angles of the wind power are analyzed and classified into deterministic and stochastic terms. The estimated PSD is modeled through piecewise affine functions, and trends of affine functions are encapsulated in multiple linear regression models. As the variability of wind power decreases, the slope of the third affine function also decreases. Furthermore, the changes of the third slope are a function of total capacity, number of wind farms, standard deviation, and mean. Therefore, the PSD of future wind power can be estimated from multiple linear regression models, and the variability of wind power can be quantified through the estimated PSD.

Short-term prediction of wind farm output using the recurrent quadratic volterra model

This paper presents a way to use the recurrent quadratic Volterra system to forecast the wind power output. The recurrent quadratic Volterra system is a second-order polynomial equation that uses the output data as feedback recursively. The Volterra system is extracted from the weights of the Recurrent Neural Network. During this process, three innovative techniques are used. In order to make Volterra kernels from the combination of weights, the activation function is approximated to the high-order polynomial function by using the Lagrangian interpolation. Furthermore, the memory of the Volterra system is also identified using the Partial Autocorrelation Function. After building the Volterra system, the 15 and 30-minutes ahead of wind power output is forecasted with confidence intervals at the 95% confidence level. The confidence interval is calculated using the multi-linear regression techniques. The stability of the recurrent Volterra system is also considered by the heuristic method.

Wind power scenario generation for stochastic wind power generation and transmission expansion planning

Wind power scenarios from individual wind farms are synthesized using given sample paths of wind power from these same wind farms through the generalized dynamic factor model (GDFM), where wind power scenarios are represented as the multiplication of a fixed polynomial matrix and time varying dynamic shocks, which are white noise. The stochastic structure among wind power outputs is nested in the polynomial matrix, but inherent wind power intermittency is nested in the dynamic shocks. We can generate an arbitrary number of scenarios by changing the dynamic shocks. Since the dimension of dynamic shocks is generally less than the dimension of wind farms, we can reduce the number of random variables. Synthesized sample paths of wind power from candidate wind farms might be used as scenarios to plan the wind power generation and transmission expansion.

Probabilistic wind power forecasting based on the laplace distribution and golden search

The point forecast of wind power and its error distribution are estimated under the assumption that the error distribution follows known distributions in a closed form. The point forecast is estimated via the gradient boosting machine (GBM). The mean of the error distribution is assumed to be zero, and the standard deviation (STD) of the error distribution is found to minimize the sum of the mismatches between the quantiles of the error distribution and the actual target value. The mismatch is measured by the pinball loss function, and the optimal STD is found through the golden section search method. The performance of our proposed algorithm is verified by using the numerical weather prediction (NWP) and wind power data from the 2014 Global Energy Forecasting Competition (GEFCom). Our current benchmark ranking is second according to the published competition result of 14th week.

Synthesis of sample paths of wind power through factor analysis & cluster analysis

The purpose of this paper is to synthesize sample paths of wind power from a wind farm in a certain location by considering the wind power variability in that location. In order to achieve this purpose, factors driving wind power variability are extracted from wind power time series through factor analysis. Then, wind farms with similar factor loadings are clustered through an agglomerative hierarchical cluster analysis. In those two processes, the number of factors and clusters are decided heuristically. The seasonal variation of wind farm clusters is also analyzed, and the relative variability among wind farms is plotted in a dendrogram of wind farm clusters. The factor loadings and statistical characteristics of wind power in the selected cluster are used to synthesize the future sample path of a wind farm in that cluster. Finally, the synthesized wind power is verified by testing whether the distribution of the minute by minute difference between wind power data follows the Laplace distribution.

CPS1-Compliant Regulation Using a PSD Analysis of Wind Expansion in a Single Balancing Authority

The integration of wind power has imposed new requirements for regulation reserves. For ensuring adequate regulation reserves, wind power variability and uncertainty must be estimated, and system frequency must comply with frequency performance standards. This work determines regulation capacity and ramping capability requirements as a function of wind-installed capacity to comply with CPS1, the North American Electric Reliability Corporation (NERC) Control Performance Standard 1 for the case of a single balancing authority. A simulation of the Electric Reliability Council of Texas (ERCOT) will be presented to examine the performance of the formulation.