Backend Architecture and Functionality:

 The TTS tool is integrated into the Pani Zed digital solution. In general, data from your SCADA system is transferred to Pani Zed and saved. From there, various calculations are done to preprocess and transform the data. This creates the rest of the KPIs visible in the product. The TTS tool intakes these general calculations, sensors, and plant configuration data and provides an updated forecast of multiple parameters once a day.

Forecasting Models and Formulas:

The model is a proprietary combination of a physical-chemical model and machine learning methodologies. The membrane’s current state, and rate of degradation are modelled as a function of water processed, not time. Assumptions are made that the recent influent conditions and operating methodology will continue. That allows us to forecast the membrane state and from that all surrounding forecasts. Water quality is a function of the membrane state.

Reliability and Performance:

When Pani sets up this tool, we do a complete backtest on all historical data available. This means that on every date in the past that we have data for, we simulate the model’s performance and then compare that to the real historical values. We set a threshold of average acceptable forecast error, and then adjust the length of the forecast to meet that accuracy.

Optimization and Best Practices:

• The model trains on old data and recent data. It requires sufficient data to produce a forecast. If too much data is missing, it will not produce a forecast. This is a primary concern for Koyambehu as there has not been much historical data uploaded. If we wait for more data to be collected by Pani, the model will eventually start producing forecasts
• Cleaning and replacement dates are key inputs to the model. It’s expected that these are entered as special logs in the Data page in the Pani-configured workflow for this purpose. If these are not entered, the model will perform poorly
• The model assumes stable train operation. If train setpoints are being rapidly adjusted, the model may not produce a forecast and will need 4-5 days of stable operation to produce a forecast
• The limits that the model warns at are adjustable by the end users to best match plant operating practices. Please adjust the limits using the settings pane as necessary
• The costs that the model uses to make economic calculations are adjustable by the end users. Please adjust the costs using the settings pane as necessary

We expect that someone checks the forecasting tool once a day at maximum, to see the updated results. The tool will give you advance warning of the membrane parameters crossing limits. Once a parameter is forecasted to cross a limit, the tool will change color. The user should double-check that the limit is correct, and verify data against plant conditions. From there, the user should have plenty of advance warning to schedule a clean close to the crossing date. Once the clean has been completed, it needs to be logged in the Data page.