Sooyeon Kim

We synthesize future scenarios of total wind power generation (WPG) of future wind farms by considering the smoothing effect, in which mid-term variability fades out, but periodicity and short-term variability remain as more geographically distributed wind farms are aggregated. First, we decompose the WPG into a periodic component (PC) and non-periodic component (NC). We design the future PC by modifying the standard deviation and mean of the current PC. Then, we assume that the NC can be decomposed into a variable component (VC), which is a source of mid-term variability, and an uncertain component (UC), which is a source of short-term variability. We separate the VC and UC in the frequency domain. We design the VC by modifying the slope of power spectral density (PSD) to represent the smoothing effect and the UC by synthesizing the PSD distribution to represent randomness. Based on our scenarios, we develop a new flexibility index and a novel recursive algorithm to calculate the required amount of future flexible resources and the value of the smoothing effect to maintain the flexible resource adequacy. Finally, we verify the smoothing effect and the amount of flexible resources by simulating the IEEE test network based on actual WPG in Texas and Australia.

We propose an innovative electricity demand forecasting framework based on three model-selection techniques to maximize the forecasting accuracy. In the framework, we forecast the day-ahead electricity demand every 15 minutes based on temperature and cloud data, which are sampled from 35 weather stations. We develop three progressive techniques for selecting the model and data. First, using a stagewise forward technique, we select highly-correlated weather stations and group the best combination of selected stations. Second, using a piecewise series technique, we select the best performing forecasting machine every hour by comparing the forecasting accuracy of four forecasting machines. Third, we develop a pairwise mapping technique to combine two tandem forecasting models at the smaller sampling interval when the sampling intervals of weather and demand data differ. We verify that the framework based on three selection techniques results in higher forecasting accuracy using data from the 2018 RTE demand forecasting competition held in France.

A delivery service using unmanned aerial vehicles (UAVs) has potential as a future business opportunity, due to its speed, safety and low-environmental impact. To operate a UAV delivery network, a management system is required to optimize UAV delivery routes. Therefore, we create a routing algorithm to find optimal round-trip routes for UAVs, which deliver goods from depots to customers. Optimal routes per UAV are determined by minimizing delivery distances considering the maximum range and loading capacity of the UAV. In order to accomplish this, we propose an algorithm with four steps. First, we build a virtual network to describe the realistic environment that UAVs would encounter during operation. Second, we determine the optimal number of in-service UAVs per depot. Third, we eliminate subtours, which are infeasible routes, using flow variables part of the constraints. Fourth, we allocate UAVs to customers minimizing delivery distances from depots to customers. In this process, we allow multiple UAVs to deliver goods to one customer at the same time. Finally, we verify that our algorithm can determine the number of UAVs in service per depot, round-trip routes for UAVs, and allocate UAVs to customers to deliver at the minimum cost.

We propose a model predictive control-based optimal offer and operation strategy for a photovoltaic (PV) farm consisting of PV panels and dual energy storage systems (ESS)s to maximize profits in the energy and regulation markets. Although a PV farm owner can better respond to regulation signals with an ESS, it cannot continuously respond to unidirectional regulation signals since the ESS is limited in size. Furthermore, the lifespan of the ESS might be reduced by alternating between charging and discharging because of regulation signals that fluctuate often. To improve the response, we use dual ESSs with separate signals: one small, fast ESS to respond to fluctuations, and one large, slow ESS to respond to unidirectional signals. We decompose the regulation signal into two signals: one for fast and one for slow frequencies based on power spectral density (PSD) analysis. We determine the optimal operation for dual ESSs to maximize the profit through a closed-loop model predictive control (MPC). We can also increase the lifespan of each ESS by limiting their state of charge (SOC) levels through the optimization constraints obtained from the PSD analysis so that the dual ESSs can operate a larger number of cycles. We verify our approach by adjusting the day-ahead (DA) market schedule in the real-time (RT) using the most recently predicted signals and errors between the DA and RT market situations. We show that our strategy incurs lower penalties by tracking the actual regulation signal in the RT market better than conventional approaches based on open-loop controls.

We propose a new method that improves the prediction accuracy of wind power generation by using two machine learning algorithms, support vector machine (SVM) and gradient boosting machine (GBM). We participate in the wind power forecasting competition held by KPX to verify the performance. First, we construct individual models in parallel using the data only at the corresponding target time since the data quality of weather data decreases as the target time increases. Second, we use the ensemble method by using two machine learning algorithms, SVM and GBM. Third, we extend the wind power generation data by interpolation to reduce the variation and estimate actual wind power generation. Fourth, we reconstruct the extended wind power generation data to prevent from converging to the average value and. We describe characteristics of stepwise model and present each result with normalized mean absolute error.

We predict electricity demand in a very high accuracy with various machine learning algorithms by designing optimal combinations of weather data, stations, and years. First, we create a novel forecasting structure by taking advantage of stair-wise structure of predicted weather data to increase the forecasting accuracy. Second, we test various machine learning algorithms : ridge regression, support vector machine, and gradient boosting machine (GBM). We find that the GBM has the best forecasting accuracy. Third, we select weather data, years of data, and weather stations based on root mean square errors. We also choose the weather stations by identifying a correlation between weather data and demand to compare two station-selecting methods. Accordingly, we increase the computation speed and forecasting accuracy by filtering out unrelated data. Finally, we verify our proposed algorithm by participating in the 2018 world electricity demand forecasting competition held in France.