Short-Term Load Forecasting Using System-Type Neural Network Architecture

This paper presents a methodology for short-term load forecasting using a system-type neural network based on semigroup theory. A technique referred to as algebraic decomposition is proposed for the modeling of electric power load demand in terms of the coefficient vector and the basis vector, and a new learning algorithm based on semigroup theory is put forward for extrapolation of the coefficient vector. Due to the non-stationary attribute of the load, the actual load is preprocessed by regression to become better correlated to daily time and temperatures. A rearrangement method based on the hourly temperature is developed to solve the problem of the roughness of the coefficient vector. With the algebraic decomposition of the rearranged regression load, a much smoother coefficient curve can be obtained. Based on the smoothness, interpolation and extrapolation can be achieved for each hour using the historical hourly temperatures and the hourly temperature forecast. The interpolated or extrapolated coefficient vector is recombined with the basis vector for each hour, and the recombined hourly load are grouped to form the final load forecast of the target day. A moving window slides through the whole year to perform the day-ahead load forecasting. Load data from New England Independent System Operator (ISO) is used to verify the capability of the proposed approach.