Reservoir inflow forecasting using neural networks

In utilities using a mixture of hydroelectric and nonhydroelectric power, the economics of the hydroelectric plants depend upon the reservoir height and the inflow into the reservoir for several days into the future. Accurate forecasts of reservoir inflow allow the utility to feed proper amounts of fuel to individual plants, and to economically allocate the load between various non-hydroelectric plants. Neural networks provide an attractive technology for inflow forecasting, because of (1) their success in power load forecasting, and (2) because of the availability of relevant measurements, including historical data. However, there are several problems that must be overcome before neutral network load forecasting systems can be successfully applied to inflow forecasting. First, forecast inflows cannot be used as inputs to inflow forecasting networks, since that has been observed to lead to unstable behavior. Second, the day of the week is of no importance in inflow forecasting. Third, too many networks are required if a separate network is used to forecast each hour in the future, over the period of several days. Fourth, each network requires hundreds of inputs, since the authors use hourly inputs over the past week in inflow forecasting. Fifth, forecasting systems using multiple networks often havemore » too little training data per network. In this paper the authors describe a system which avoids these problems. First they describe the features and the desired outputs related to their system. They also introduce some notation. Then they discuss compression by Karhunen-Loeve transform and its application in the reservoir inflow forecasting system. Next they describe the neutral network used and the training algorithm adopted for the system. Lastly they discuss some forecasts produced by the system.« less