The input raw material to a waste-water treatment plant exhibits large, and generally poorly quantified, variations with time. In particular, rainfall run-off can cause gross overloading of the treatment processes of the plant. For a proper operational control of the plant, and hence the quality of the receiving river's water, it would be extremely useful to have advance (short-term) estimates of the effluent flow from the sewer network, i.e. the influent to the plant. This paper studies the feasibility of using an on-line adaptive predictor in such a capacity. The procedure is divided into two steps : (i) the parameters of a multiple input/single output time-series model are recursively estimated at each time-step by the method of least squares ; (ii) a forecast of the plant influent flow is then made on the basis of the newly updated prediction model. Results are presented for data from a treatment plant in Stockholm, Sweden. These demonstrate the adaptability of the predictor to unknown changes in the process dynamics when no information is assumed to be available for rainfall events occurring over the urban land surface,