Data driven insights that provide value at every step of the way from process design, operating water treatment plants and managing sewer networks.

Currently, EVS Water operates only in Amazon Web Services (AWS) infrastructure, as it is tightly coupled to AWS capabilities for machine learning and data science. There are some possibilities for hosting some data locally, but these are highly dependent on specific project requirements, would be considered and would need to be discussed with our data management team.

EVS Water uses recurrent neural networks (RNN) powered by TensorFlow and Keras to calibrate models around a deterministic baseline. These tools have been chosen as they are widely accepted frameworks for applying RNN to challenges requiring identification of patterns in complex data sets. These frameworks are production-focused and integrated into the source code to enable accurate forecasts for EVS Water applications and allow for relatively rapid implementation and processing times. Input features to the RNN may include Feed/Coagulation pH, DOC, UV254, Temperature, Feed/Lamella Overflow Turbidity and coagulant dose, using both existing and a vendor installed instruments.

Sorption Capacity used in the deterministic model is also further calibrated by this RNN model to site specific conditions. The detailed design of the RNN models is developed as a result of close engagement with the customer and site personnel and depends on the design of the plant and available monitoring data.

We use a procedure for forecasting time series data based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality, plus holiday effects. It works best with time series that have strong seasonal effects and several seasons of historical data. It is also robust to missing data and shifts in the trend, and typically handles outliers well.

The deterministic baseline for coagulation recommendations are based on a number of key models including a modified version of the Edwards (1997) Langmuir based coagulation model which is developed specifically for representing enhanced coagulation using aluminium and ferric salt based coagulants, considered the most accurate of the available models for DOC removal during coagulation (Tseng and Edwards, 1999).

The key inputs for this model are feed DOC [mg/L] and feed/coagulation pH which lead to a prediction of coagulant dose [mg/L], based on a recent inorganic analysis or online alkalinity measurement if available.

PIDs, PFDs, monitoring points, representative monitoring data, dosing locations, equipment datasheets will be required to be shared at the start of the project. Data for nominated monitoring points will need to be shared at least daily during the operational phase of the project with protocols agreed at the start of the project.

To facilitate an evaluation of financial benefits, average unit cost of chemicals, electricity consumption, solid waste disposal and labour cost for the plant will be requested at the start of the project.