Dr Yifan Zhang1, Mr Peter Fitch2, Dr Peter Thorburn1
1CSIRO, St Lucia, Australia,
2CSIRO, Canberra, Australia
Water quality is an important issue because of its effects on human health and aquatic ecosystems. An understanding of the long term trends in water quality is extremely important for scheduling water quality management activities.
Data-driven models have gained much attention for predicting nonlinear time series in hydrological modelling, yet predicting long term water quality change is still a big challenge. Firstly, most of these studies are restricted to predicting water quality in one short upcoming time step. In this approach, hourly/daily predicative models only predict water quality in the next one hour/day, and provide no information on the longer-term trends in water quality. Secondly, while some data-driven models can predict monthly or yearly water quality changes, they either use or resample the data with monthly or yearly time interval. Therefore, the ‘long term’ prediction still follows the one time step idea and has the same single prediction issue.
We proposed a deep learning based method that we call Deep Regressor Chain (DRC) to overcome the above issues. DRC connects multiple recurrent neural network (RNN) models in order. The 1st RNN uses N numbers of time series data and predicts at time step N+1. The 2nd RNN combines the previous RNN’s inputs and prediction together as the new input to predict at time step N+2. Followed by this hybrid strategy, DRC can predict long term water quality in N+m upcoming time steps at once by integrating all the previous predictions and no extra water quality data resampling work is needed.
Yi-Fan Zhang is a Postdoctoral fellow in Agriculture & Food, CSIRO. He received a PhD in data science from Queensland University of Technology in 2016. His work focuses on deep learning for agriculture decision making and management, with an emphasis on water quality time series modelling and forecasting.