COMPARISON OF 2 MODELS FOR REAL-TIME DAILY FORECASTS OF INFLOWS

This study presents a comparison of the performances of two models use d for real-time forecasting of daily inflows to the Chute-du-Diable an d the Lac-Saint-Jean reservoirs, and the daily flows measured on the M istassibi River basin. The three drainage basins are located in the Sa guenay-Lac-Saint-Jean water system. The first model, a conceptual one, is a global deterministic model that is currently being used by Alcan (Aluminum Company of Canada) to predict daily flows in real time. The second model, which forms the primary focus of this study, is based o n the structure of models commonly known as ''black-box'' models, a ge neralized formulation of the autoregressive moving average models with exogeneous variable of Box and Jenkins (ARMAX). The Kalman filter is coupled with this model to enable day-to-day adjustment of estimated p arameters. Autocorrelation and cross-correlation analyses on the data have made it possible to establish the preliminary structure of the bl ack-box model, that is, one per basin. The final structure was chosen following a sensitivity analysis on the parameters. The models retaine d are all ARMAX models, the statistical behavior of the residuals havi ng demonstrated their adequacy. Comparison of these ARMAX models with Kalman filter and the deterministic model have led to the following co nclusion: the ARMAX models with the Kalman filter are superior to the deterministic model for daily prediction in real time within a horizon of 2 days. For a 3-day horizon, the models are, for practical purpose s, equivalent. For a horizon of 4 days or more, the deterministic mode l is superior to the ARMAX models with Kalman filter.