Enhancing accuracy of autoregressive time series forecasting with input selection and wavelet transformation

Autoregressive time series forecasting is common in different areas within water resources, which include hydrology, ecology, and the environment. Simple forecasting models such as linear regression have the advantage of fast runtime, which is attractive for real-time forecasting. However, their forecasting performance might not be acceptable when a non-linear relationship exists between model inputs and output, which necessitates the use of more sophisticated forecasting models such as artificial neural networks. This study investigates the performance and potential of a hybrid pre-processing technique to enhance the forecasting accuracy of two commonly used neural network models (feed-forward and layered recurrent neural network models) and a multiple linear regression model. The hybrid technique is a combination of significant input variable selection (using partial linear correlation) for reducing the dimensionality of the input data and input data transformation using discrete wavelet transform for decomposing the input time series into low and high frequency components. Two case study forecasting applications, namely, monthly inflow forecasting for a lake in Victoria (Australia) and weekly algal bloom predictions at a bay in Hong Kong were used to assess the forecasting ability of the models when used in conjunction with the hybrid technique. Results demonstrated that the hybrid technique can significantly improve the forecasting performance of all the models used.