Electric Power Systems Research 59 (2001) 121 – 129 An adaptive neural-wavelet model for short term load forecasting Bai-Ling Zhang a , Zhao-Yang Dong b, * a Kent Ridge Digital Labs (KRDL), 21, Heng Mui Keng Terrace, Singapore 119 613, Singapore b School of Computer Science and Electrical Engineering, Uniersity of Queensland, St. Lucia, QLD 4072, Australia Received 10 January 2001; received in revised form 18 April 2001; accepted 7 June 2001 Abstract This paper proposed a novel model for short term load forecast in the competitive electricity market. The prior electricity demand data are treated as time series. The forecast model is based on wavelet multi-resolution decomposition by autocorrelation shell representation and neural networks (multilayer perceptrons, or MLPs) modeling of wavelet coefficients. To minimize the influence of noisy low level coefficients, we applied the practical Bayesian method Automatic Relevance Determination (ARD) model to choose the size of MLPs, which are then trained to provide forecasts. The individual wavelet domain forecasts are recombined to form the accurate overall forecast. The proposed method is tested using Queensland electricity demand data from the Australian National Electricity Market. © 2001 Elsevier Science B.V. All rights reserved. Keywords: Neural network; Time series; Adaptive learning; Wavelet; Load forecast www.elsevier.com/locate/epsr 1. Introduction Electricity industry is undergoing deregulation in many countries. The aim of deregulation is to free the customers of their choices of electricity supply, and to maximize the social benefit of the society. Driven by the market, many systems are being pushed in stressed situations close to their security margins. The system security is becoming more and more important in the electricity industry. However, many of the security problems are caused by various contingencies out of system operation and control. One of the essential factors to manage and control such contingencies is by proper system operational planning based on the load forecast. The aim of short term load forecast is to predict future electricity demands based, usually, on historical data and predicted weather conditions [1 – 3]. Then such important operational and management de- cisions as generation scheduling, maintenance schedul- ing, ancillary services scheduling and risk management can be carried out based on the predicted electricity demands. Traditionally, short term load forecast techniques use statistical models, which include peak load models and load shape models [2,3]. The load shape models relay on time series analysis techniques. The Autoregressive moving average (ARMA) model is among the most popular ones of dynamic load shape models. Forecast implementations in Energy Management System (EMS) system based on these linear regression forecast meth- ods have been practiced by utilities such as PG&E [2]. However, these models and techniques are basically linear methods, which have limited ability to capture nonlinearities in the load series. Artificial intelligence methods for forecasting have shown ability to give better performance in dealing with the nonlinearity and other difficulties in modeling of the time series. Artificial neural networks (ANN) have been practiced recently in the area of time-series forecasting due to their flexibilities in data modeling [3,4]. In this paper, we study neural network models combined with wavelet transformed data and show how useful infor- mation can be captured on various time scales. These strategies approximate a time-series at different levels of resolution using multiresolution decomposition. An auto-correlation shell representation technique is em- ployed to reconstruct singles after wavelet decomposi- tion. With the help of this technique, a time series can be expressed as an additive combination of the wavelet coefficients at different resolution levels. These tech- * Corresponding author. 0378-7796/01/$ - see front matter © 2001 Elsevier Science B.V. All rights reserved. PII:S0378-7796(01)00138-9