Time‐Series Modeling for Long‐Range Stream‐Flow Forecasting

Abstract

Currently used methods for long‐range water‐supply forecasting are compared with statistical time‐series tools, such as seasonal auto‐regressive integrated moving‐average modeling. Evaluation of several theoretical models under a range of flow conditions provided insight into development of a technique using engineering knowledge and experience to improve the quality of forecasts. Rules governing model selection are developed from analysis of forecast residuals within a sensitivity analysis. Context‐sensitive model selection provided a means of improving forecast accuracy. Increased confidence in the optimal forecasted operating and planning policies are consequences of improved forecasts. The modeling tools provide the means of evaluating the performance of long‐range monthly probabilistic stream‐flow forecasts at Manitoba Hydro. Manitoba Hydro is a large utility that operates a multireservoir electric‐power generation system. The needs and priorities of the system demand forecasts up to 1 year in advance for planning budgets and release policies.