Electricity industry is important for our society and the nation's economic strength, and this dissertation includes two important topics in the industry: electricity market auctioning, and short-term load forecasting. In the first part of this dissertation, an optimization algorithm to solve the payment cost minimization auction for deregulated electricity markets is presented. In the second part, a novel short-term load forecasting (i.e., forecasting the next day's load) method is presented.
Deregulated electricity markets in the U.S. currently use an auction mechanism that minimizes total bid cost to select bids, but determine payments based on market clearing prices (MCPs). Under this setup, consumer payments could be considerably higher than the minimized cost obtained from auctions. This gives rise to "payment cost minimization," an alternative auction mechanism that directly minimizes consumer payments. We have previously presented an augmented Lagrangian and surrogate optimization framework to solve the payment cost minimization problem for energy. The first part of this dissertation extends this framework to solve payment cost minimization with simultaneous optimization of energy and ancillary services. The problem is difficult since MCP, unit capacity constraints and energy/ancillary service bids are related in a complicated way. To solve the problem, unit capacity constraints are relaxed, and unit and demand bid subproblems are formed within the augmented Lagrangian and surrogate optimization framework. A unit subproblem is solved by using dynamic programming (DP), with its energy and ancillary service levels separately optimized for each DP state. Numerical results demonstrate the quality and scalability of this method, and show that simultaneous optimization yields significant savings as compared to sequential optimization.
Short term load forecasting is essential not only for reliable power system operations, but also important for market participants. It is, however, difficult and challenging. One of the widely used forecasting methods is neural network. In view of the complicated features of load, the standard neural network method sometimes fails to provide accurate forecasts. The second part of this dissertation presents a novel similar-day based wavelet neural network method for improving short-term load forecasting. A similar day technique is used to select good inputs. Wavelet is used to decompose the load into a low frequency component and a high frequency component. Two separate neural networks are then used to capture the features within the two frequency components. Numerical testing shows that this method significantly improves prediction accuracy. This method is being implemented by ISO New England for their daily use.
