Adaptive forecast of multi-month lake level elevations
The early and precise prediction of the water levels in lakes is a major concern for public authorities. Such predictions describe the evolution of the water levels and are essential for appropriate flood control measures. In this paper, a new ARMAX-type model is developed to predict, months in advance, the monthly fluctuations of the water level of Lake Erie. The predictive variables used in the model are the past monthly water levels of Lakes Erie, Superior, and Michigan–Huron along with the estimated response times between water flow entries and exits. Two scenarios are compared. The first scenario is based on the ordinary least squares (OLS) technique in order to identify the parameters of the ARMAX-type model, to filter measurement and model noises, using the ARMAX Kalman predictor (AKP), and to optimize predictions. In this scenario, the model parameters remain unchanged throughout the simulation. The second scenario is based on the mutually interactive state parameter (MISP) technique in order to readjust the parameters of the model at each time step and to filter measurement and modelling noises through the Kalman predictor. In this scenario, the parameters of the model change with time. The analysis shows that the MISP–AKP framework has a slightly higher Nash coefficient than the OLS–AKP framework for the first month. In subsequent months, however, the quality of the predictions based on the OLS–AKP technique improves significantly. This observation also applies to the persistence and extrapolation coefficients as well as to the sample autocorrelation functions for the residuals of the Lake Erie water level forecast. It was therefore decided to apply the MISP–AKP technique to obtain the first prediction of the Lake Erie level, and the OLS–AKP technique to compute subsequent predictions. Key words: adaptive, forecast, Kalman's filter, lake levels, MISP algorithm, Great Lakes.