Water Temperature Forecasting in a Small Stream Using Neural Networks with a Bayesian Regularization Technique

摘要

Physical processes influencing water temperature in a river are highly complex and uncertain, which makes it difficult to capture them in some form of deterministic model. Accurate forecasting of water temperature in a river is important, as it has implications on the quality of water and the lives that depend on it. Here we develop a model of forecasting which allows estimation and forecasting of water temperature at short and middle term, It intends to forecast the water temperature of days (t+i, i=1,2…), t is the current time. Due the strong dependence between water temperature at the current time and those for the past, the projected model builds easily itself, by investigating, for each stage of forecasting, the function relating input and output relationships. For this, a multi-step-ahead forecasting model based on the neural networks with the Bayesian regularization technique, is formulated for establishing linkages between water temperature and influencing variables. The results show that the elaborated model is robust and reliable and gives good results. It allows us to forecast the water temperature with high success. To test the ability of the model for the prediction, the observed data of the average daily water temperature during a period of five years (1998-2002) is considered for analysis. The first three years serve for the training and the remaining for the test. The model produced a standard coefficient R about 98, while the standard deviation s does not exceeds 0.6°C. We noticed there are a few cases presenting an error between 1 and 1.5°C (On average three cases for all steps of forecasting).