8-Yang-Wastewater-Treatment-Optimization-Using-Data-Driven-AI-ML-Models.pdf
- 1. Fenghua Yang, P.E. BCEE Senior Environment Research Scientist | MWRD
- 2. Outline • Data-Driven ML vs. Mechanistic Models • Case Studies Using Data-Driven ML Models • Data management and security Consideration • Challenges and Opportunities
- 3. Mechanistic vs. Data Driven Models in Digital Twin Physical System/Process Mechanistic Models - Biochemical process modeling (GPS- X, BioWin etc.) - Hydraulic modeling (visual hydraulics, etc.) Data Management - Data storage, Data integration, Data analytics etc. Visualization - Data & Trends, KPI’s, Prediction, etc Data Driven Models (AI/ML): - Classical machine learning (RF, SVM etc.) - Neural Networks (RNN, etc.) Digital Twin Plant
- 4. Mechanistic vs. Data Driven Models in Digital Twin Mechanistic Models - Biochemical process modeling (GPS-X, BioWin etc.) - Hydraulic modeling (visual hydraulics, etc.) Data Management - Data storage, Data integration, Data management Visualization - Data & Trends, KPI’s, Prediction Data Driven Models (AI/ML): - Data & Trend - Process optimization - Prediction Data + Knowledge
- 5. Data Drive AI/ML in WastewaterTreatment 1. Advanced Control for Swing Zone Operation to Balance and Maximize Enhanced Biological Phosphorus Removal (EBPR) and Ammonia Removal - Proof-of-Concept Study (2018) 2. Optimize Disinfection using AI/ML – Project Planning 3. AI/ML in ABAC – Project Planning 4. Machine Learning Application for Headworks Odor and Corrosion Control – Full Scale Test and Implementation (2019- 2022)
- 6. Case Study 1 - Advanced Control for Swing Zone Operation (Proof-of-Concept ) • Teamed up with inCTRL Solutions • Project goal: Advanced control for swing zone operation to balance and maximize BioP and ammonia removal • Data: 2.5 years routine data: online instrument data (15-min or 1-hr interval) and Lab data (daily) • Method: Multivariate statistical methods were used to develop a data driven predictive model (soft sensor) • Status: Proof of Concept
- 7. Case Study 1 - Multivariate Statistical Methods
- 8. Case Study 2 – Optimizing Disinfection using AI/ML Models (Planning) • Current Challenges: • Flow can vary from 20 - 120 mgd • Flow monitoring at headworks - difficult to utilize ‘real-time’ flow • Residual analyzers/testing method limitations • Chlorine Analyzer Detectable limit of 0.02 mg/l and no in-stream fecal monitoring DISINFECTION AT KIRIE WRP • 270k gallons of NaOCL and 16k gallons of sodium bisulfite used annually POTENTIAL EXPANSION TO OTHER DISTRICT WRPS • District uses approx. 1.8 MG of sodium hypochlorite ($1.3M) per year • Potential Benefits:
- 9. IWS potential Implementations in WRP – Optimization of Ammonia Based Aeration Control (ABAC) Current feedback cascade ABAC in Stickney provides potential of 36% air savings • Effluent ammonia spikes were observed due to the delay and limitation of air supply system – risk of permit violence • Significant amount of time for sensor maintenance • Significant amount of time for data QA/QC • Interrupt ABAC operation due to sensors out of service Benefits of adding ML to the feedback cascade ABAC • Add data-based feedforward control to reduce effluent ammonia spikes – performance improvements • Soft sensor to reduce sensor maintenance, reduce time for data QA/QC, and provide continued ABAC control when physical sensors out of service
- 10. IWS potential Implementations – Soft Sensors • Physical sensor/ soft sensor cross validation • Provide automatic fault detection of physical sensors for outliners, error data (for sensors out in the field, or sensors tends to drift) • Filling the missing or bad data and can be used as backup signal for control purpose when physical sensors goes offline. • Identify changes in process or operational issues • For parameters that online instrument is currently not available or costly to buy and maintain (VFA, Ortho-P, etc)
- 11. IWS potential Implementations in Wastewater Industry – Process optimizer • Can be implemented in any process that has sufficient data and needs operational control (screen cleaning cycle, RAS control, WAS control, aeration control, chemical dose, filter backwash, UV disinfection, polymer dose…) • Use domain knowledge to identify and prioritize implementation to achieve optimal chemical/energy savings and performance improvements
- 12. Summary • When should I consider ML • When mathematic is not available and you have sufficient data • When you want to do prediction, process optimization, real time decision making • What expertise needed • Process engineer/operation staff • Data analytics Unit Process Input Output Energy Chemicals Others data data d a t a d a t a d a t a Modeling Unit Process
- 13. Case Study - Background NaOCl dosing for Headwork Odor/Corrosion Control EPBR implementatio n in 2015 Suspension of NaOCl dosing Resume NaOCl dosing? How ? 87% overdosing 5% under dosing Team up with Data Analytics from ISU for the 2019 IWS Challenge
- 14. Using ORP for Controlling NaOCl dosing? 0 50 100 150 200 250 300 -500 -300 -100 100 300 3/7/2019 3/14/2019 3/21/2019 3/28/2019 4/4/2019 4/11/2019 4/18/2019 4/25/2019 5/2/2019 5/9/2019 5/16/2019 5/23/2019 5/30/2019 6/6/2019 6/13/2019 6/20/2019 6/27/2019 7/4/2019 7/11/2019 7/18/2019 7/25/2019 8/1/2019 8/8/2019 8/15/2019 8/22/2019 H2S (ppm) ORP (mV) ORP (mV) 0 mV H2S (ppm) 85% overdosing 3% underdosing
- 15. Using Online H2S Meter for Controlling NaOCl dosing? • Permeant online H2S to provide 4-20mA signal to control the dosing • Target H2S in flow distribution box headspace < 5ppm • Typical feedback control • Concerns • Online H2S sensor reliability • Impact to VFA cannot be evaluated • Potential overdosing – Impact downstream treatment (nitrification and bioP) • Responsive – Lag time • No prediction
- 16. Some Facts of using NaOCl to Remove H2S Fundamentals • H2S (water phase) level is impacted by many factors: wastewater characteristic, ORP, other organics etc • NaOCl reacts readily with H2S, but also reacts with any other oxidizable material, VFAs. Reaction rates depend on chemical energy level of each reaction and other factors (pH, temp, concentration etc) and is hard to determine • H2S (air phase) is problematic. It is impacted by many factors: H2S level in the water phase, pH, temp, turbulence etc. • Lack of mechanistic modeling approaches • Excess chlorine often used to force the reaction for H2S (water phase) removal.
- 17. Water Quality Analysis Data Data QA/QC and AI Models Predicted Water Quality Analysis Data Data QA/QC and AI Models Predicted H2S class Predicted VFAs class Predicted optimal NaOCl dosing Data QA/QC and AI Models Online Instrument Data Wastewater flow, target H2S, and target VFAs Module 1 Module 2 Module 3 Precipitation Data 17 years of Influent historical data: BOD, Ammonia, TKN, TP, SO4, TSS, TS 6 months online instrument data (ORP, Temp, pH, Flow, TARP Elevation, and CUP pumping 10 weeks of NaOCl dose-response data
- 18. Footer put title here. – 18 Module 1 Module 2 Module 3 Module Goal Predict influent wastewater characteristics to use in Module 2 (TS, SS, TP, NH3-N, SO4, BOD5, Org- N, and TKN) Predict H2S and VFAs to use in Module 3 Predict chemical dosage of sodium hypochlorite Available Dataset 17 years of influent historical data which had good quality Six months of online instrument data which were insufficient and imbalanced 10 weeks dose response data which were insufficient and lack in variation Feature Variables TS, SS, TP, NH3-N, SO4, BOD5, Org- N, TKN, and precipitation Model 1 output data (predicted TS, SS, TP, NH3-N, SO4, BOD5, Org-N, and TKN); online instrument data (flow, ORP, pH, wastewater temperature, tunnel pumping, and tunnel elevation) Flow, Model 2 output (predicted H2S and VFAs), target H2S, and target VFAs Data Preprocessing Method data filling Quantization, Quantization, oversampling oversampling, data filling Models Tested in each Module RNN with LSTM, ARIMA, RF, XGBoost, SVM RF, SVM RF, SVM Model Chosen for Module Regression using deep learning model (RNN with LSTM) Classification using classical machine learning model. RF was the best at predicting VFAs while SVM was the best at predicting Classification using classical machine learning model (RF)
- 19. Module 1 Selection: • RNN (LSTM) • RF • ARIMA 0 0.1 0.2 0.3 0.4 0.5 0.6 NH3 BOD5 TS Org-N TP SS TKN SO4 MAE RNN ARIMA RF Module 2 Selection: • RF Classifier • SVM Classifier Module 3 Selection: • RF Classifier • SVM Classifier Outpu t Method Data interval Accurac y in testing set (%) H2S RF 15- minute 87.6 SVM 86.3 RF daily 85.7 SVM 97.6 VFA s RF daily 93.4 SVM 91.0 RF SVM 68.75% 89.75%
- 20. Performance: Kirie WRP Treatment Optimization • 1,000+ predictions of influent water quality since August, 2021 • Augment LIMS dataset • Extrapolation up to 7 days • Data Health Feedback • H2S, VFA prediction • NaOCl recommendation NH3 Predicted vs Observed Concentration (mg/L)
- 21. Data Preparation From IWA Meta Data Management Project. 21 RAW DATA DATA FIT FOR PURPOSE • DATA VISUALIZATION:time series, histograms, scatterplots… • DATA GROUPING: dry/wet weather, cyclic patterns... • DESCRIPTIVE STATISTICS: mean, median, min, max, standard deviation, skewness coefficient, correlations among variables... • SANITY CHECKS • OUTLIERS • MASS BALANCES • TRANSFORMATION • SCALING • MISSING DATA IMPUTATION • FILTERING AND SMOOTHING • FEATURE GENERATION • DATA STRUCTURING INITIAL REVIEW ERROR IDENTIFICATION AND CORRECTION DATA PRE-PROCESSING STEPS SELECTION AND APPLICATION
- 22. Influent Water Quality Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • Typical challenge when working within a real-time framework. • The table represents a set of timeseries for several water quality parameters which are required for Model 2 and Model 3 to function. • A recursive function used to make prediction at points in time when data is missing any water quality parameter. The function takes a trained models from Module 1, and an input dataframe of WQ values with a column for each of the models along a 24-hour time- resolution datetime index. • The following animation illustrates how the missing data is resolved with neural networks. Real-Time Data Availability Matrix: Set of timeseries for influent water quality predictions
- 23. Module 1 WQ Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • 1st date with missing WQ values: • Jan 3 • Which signals missing • TKN • Predict TKN on Jan 3
- 24. Module 1 WQ Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • 1st date with missing WQ values: • Jan 4 • Which signals missing • Total Solids, TKN • Action: • Predict TKN • Predict Total Solids
- 25. Module 1 WQ Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • 1st date with missing WQ values: • Jan 5 • Which signals missing • Total Solids • SS • P-TOT • TKN • Action: • Predict Total Solids • Predict SS • Predict P-TOT • Predict TKN
- 26. Module 1 WQ Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • 1st date with missing WQ values: • Jan 6 • Which signals missing • Total Solids • SS • BOD5 • NH3 • Org-N • P-TOT • TKN • Action: • Predict Total Solids • Predict SS • Predict BOD5 • Predict NH3 • Predict Org-N • Predict P-TOT • Predict TKN
- 27. Module 1 WQ Predictions datetim e Total Solids SS BOD5 NH3 Org-N P-TOT SO4 TKN Jan 1 Jan 2 Jan 3 Jan 4 Jan 5 Jan 6 • Last Step • No missing WQ values • Return the dataframe • Note: • Error from preceding predictions accumulates in subsequent predictions
- 28. Kirie WRP Treatment Optimization, Project Objective Objective: To integrate existing algorithms developed by Illinois State University (ISU) faculty and MWRD under previous work, into a process optimization platform for the Kirie WRP. Goal: The goal is to optimize dosing of sodium hypochlorite (NaOCl) to prevent the formation of hydrogen sulfide (H2S) without disrupting the formation of volatile fatty acids (VFAs). Figure 1: Flow process of BLU-X Treatment, data-driven process optimization.
- 29. Option 1: Signal Splitter Option 2: DIODE
- 30. Integration of data into Real Time Decision Support System 30
- 31. Major Takeaways from Pilot Project: Footer put title here. – 31 • Direct, quantifiable recommendation for hypo dosing in real time through optimization; however, warrants more LIMs data • Improve data governance and data management across the District • Gained more knowledge of the feasibility of integrating research NN models into real time using Xylem’s approach Future Additional Benefits: • Early warnings of upcoming flow events, influent characteristic, H2S/VFAs levels, and Hypo dosing recommendations to the Kirie WRP. • Using this pilot study as a demonstration of the possibilities for using AI in WRPs.
- 32. Challenges • Data • Integrate with traditional mechanistic model • Transfer learning in AI/ML
- 33. Opportunities to learn and get involved • LIFT Intelligent Water System (IWS) Challenge- 2022 • Registration deadline (Raise hand) • Plan submission • Final solution submission • WEF RICE Group - Soft sensors and machine learning possibilities and applications • IWEA/IWS Committee • Webinars • Student IWS Challenge • Funded AI/ML projects • WRF • DOE Contact Information Fenghua Yang, P.E. BCEE yangf@mwrd.org
- 34. Acknowledgements inCtrl Solution Mr. Ivan Miletic and Dr. Alex Rosenthal Illinois State University Dr. Xin Fang,Dr. Yongning Tang, MWRD Ms Thais Pluth, Mr. Matt Jurjovec, and Kirie WRP Operational Staff Xylem teams • Nick Mills, Adam Erispaha, Patrick Henthorn, Xylem’s Hydroinformatic Engineers Q&A