Cost Effective Soft Sensing for Wastewater Treatment Facilities

Wastewater treatment plants are complex, non-linear, engineered systems of physical, biological and chemical processes operating at different timescales. Sensor systems are used to monitor wastewater treatment plants in order to ensure public safety and for efficient management of the plants. However, parameters of interest for wastewater can require expensive or inaccurate sensors or may require off-site laboratory analysis. For example, ammonium is important as a prime indicator of treatment efficiency and is highly regulated in discharge water. But ammonium sensors are also expensive at over $\$ $ 10,000 (AUD) per sensor. Soft sensors are computational models that accurately estimate process variables using the measurements from few physical sensors and can offer a cost-effective substitute for expensive wastewater sensors such as ammonium. In this paper, we propose a hybrid neural network architecture for learning soft sensors for complex phenomena. Our network architecture fuses sequential modelling with Gated Recurrent Neural Network units (GRUs) to capture global trends, with Convolution Neural Network (CNN) kernels to facilitate learning of local behaviours. We demonstrate the effectiveness of our technique using real-world data from a wastewater treatment plant with two-stage high-rate anaerobic and high-rate algal treatments. Secondly, we propose a novel data preparation algorithm that enables the deep learning techniques to learn from a limited data and facilitates fair evaluation. We develop and learn a soft sensor to predict ammonium and study its generalization. Our results demonstrate fit for purpose accuracy and that the soft sensor model is able to capture complex temporal patterns of the ground truth sensor time series. Finally, we publicly release an annotated data set of a secondary wastewater treatment plant to accelerate the research in the development of soft sensors.