僦又毓盅戎右×右菀×右棕仲右o颂号悄烈-Parthenope.PDF

Neural Short-term Prediction based on Dynamics Reconstruction F. CAMASTRA and A.M. COLLA Research Development Department, Elsag Bailey - Finmeccanica S.p.A., Via G. Puccini 2, 16154 Genova, Italy E-mail: francesco.camastra@elsag.it (Received ; Accepted in nal form ) Key words: Dynamics Reconstruction, electrical load forecasting, mo del order esti- mation, Multi-Layer Perceptron, short-term prediction Abstract. In this pap er we present an application of Dynamics Reconstruction techniques to mo del order estimation. Both Grassb erger - Pro caccia and Takens metho ds were applied, yielding similar values for the correlation dimension, hence for the mo del order. Based on this mo del order, appropriately structured neural nets for short-term prediction were designed. Satisfactory exp erimental results were obtained in one-hour-ahead electrical load forecasting on a six months b enchmark from an electric utility in the USA. Abbreviations: ARMA - autoregressive moving average; BIC - Bayesian Informa- tion Criterion; MLP { Multi-Layer Perceptron 1. Intro duction In the last years Nonlinear Dynamics metho ds were applied to dif- ferent elds, namely: chaotic b enchmarks, such as Henon map, Rossler oscillator [8]; natural time series, like sunsp ots and Canadian lynx p op- ulation [7]; signal pro cessing, such as radar clutter identication [6] and nancial series [11]. In particular, techniques for dynamics reconstruc- tion are employed to obtain information on the data characteristics (determinism, p erio dicity, mo del order). However, metho ds for dynam- ics reconstruction yield reliable information only when the time series is long enough and of go o d quality, i.e. not to o noisy. In our opinion,