Implementation practice of short-term load forecasting in time series

摘要

This paper presents an implementation practice of real-time short-term load forecasting using time series model incorporated with a similar-day model. In the time series model, a comprehensive load model is developed by incorporating time series functions, nonlinear load-weather functions and a residual load function. The model parameters are estimated and updated on-line using the WRLS (Weighted Recursive Least Squares) algorithm. A variable forgetting factor (VFF) technique is included in the WRLS algorithm for improved model tracking capability and numerical performance in real-time operation. The model tracking capability and numericla performance in real-time operiton. The model parameters are recursively updated in real-time based on the newly available actual load and weather data. THe time series functions are represented by Fourier series, and the important harmonic components are identified by off-line spectrum analysis. The nonlinear load-weather functions are modeled using cubic functions, and the residual load function is represneted by an ARMA (Auto-regressive Moving Average) model. In order to improve the forecasting sccuracy and reliability under certain abnormal conditions, such as severe cold fronts in winter or heat wave in summer, a similar-day model is developed by conceptually defining a set of similarity functions based on the seasons, day types, and weather conditions. A software packag, STLF, has been developed based on the proposed models. It has demonstrated successful operation in several power utilities. Off-line testing and on-line operation has consistently shown satisfactory performance.