Technical Field

The present invention relates, in general terms, to estimating outflow parameters of a treatment plant. In particular, the present invention relates to, but is not limited to, estimating outflow parameters of the wastewater treatment plant.

Background

Biological wastewater treatment processes are widely used. They can be differentiated into anaerobic wastewater treatment, anoxic wastewater treatment and aerobic wastewater treatment.

In wastewater treatment process, aerobic biological process is usually used to reduce the strength of the wastewater. In aerobic biological process, microorganisms such as bacteria, protozoa and fungi degrade parameters such as biochemical oxygen demand (BOD) and chemical oxygen demand (COD), with dissolved oxygen as the final electron acceptor. Ammonia is also removed during aerobic wastewater treatment process, via nitrification by nitrifiers. During aerobic treatment, carbonaceous BOD are oxidized, forming carbon dioxide and biomass. Non-biodegradable organics, and possibly some unconsumed biodegradable organics, will be discharged from the aerobic biological process and may be subjected to further treatment downstream. Under normal operating conditions, aerobic wastewater treatments are reliable and robust in treating wastewater to the required quality for effluent discharge.

Aerobic wastewater treatment encompasses any biological treatment processes that is performed in aerobic conditions, including, but not limited to, an activated sludge process, trickling filter, membrane bioreactor, and moving bed bioreactor. In the aeration tank, oxygen is supplied by aerators, which then dissolve in the mixed liquor and are used by the microorganism.

Anaerobic biological processes are another type of biological wastewater treatment process that is often deployed in treatment plants. Anaerobic biological process is operated in the absence of oxygen, nitrate and nitrite. It utilizes compounds other than oxygen as the electron acceptor. This results in valuable by-products including, but not limited to, hydrogen gas, methane gas and hydrogen sulfite. The absence of oxygen as final electron acceptor results in low biomass growth, hence, lower biomass disposal compared to aerobic treatment processes. This also leads to higher operating solid retention time in anaerobic treatment processes.

An anaerobic biological process is usually deployed for treatment of high strength wastewater, reducing the strength of the wastewater prior to downstream treatment. It is also often used for phosphorus removal. In the enhanced biological phosphorus removal process, an anaerobic tank is present prior to the aeration tank. In this configuration, polyphosphate-accumulating organism will be enriched, allowing biological phosphorus removal.

Anoxic biological process are another type of biological wastewater treatment process often deployed in treatment plants. An anoxic process is operated in the absence of oxygen and in the presence of nitrate, nitrite or both. In anoxic processes, denitrification occurs. In denitrification, either nitrate or nitrite is used as the final electron acceptor, forming nitrogen gas and water. This process is useful in nitrogen content removal of wastewater.

Wastewater treatment processes are deployed in municipal and/or industrial contexts. In the industrial context, wastewater treatment can be significantly more challenging as compared to its domestic wastewater counterpart. Different industries generate different types of wastewater, with different characteristics. Different industries generate different wastewater characteristics, resulting in varying characteristic such as BOD, COD, pH and temperature. Coupled with uncertainties in operating hours and processes in industrial plants, various treatment processes may be required in order to meet discharge standards. This is especially challenging for centralized industrial wastewater biological treatment plant, where feed streams may contain wastewater from different industries.

To reduce the challenge faced by centralized industrial wastewater treatment plant, simulations are often done to optimize the plant performance and to predict the performance of the treatment process. However, in a conventional mechanistic model, the user will have to either set up lab scale system, use historical plant data or perform experimental analysis to determine the necessary information required by the model. These processes are labour intensive and time consuming.

In conventional mechanistic model simulation, average biodegradability and biomass characteristics are used to simulate performance. However, as the biodegradability of wastewater and characteristics of biomass fluctuate with time, this results in inaccurate simulation.

Another approach used in the prediction of parameters of outflow from aerobic biological processes is the use of machine learning. Machine learning recognizes patterns between the various inflow parameter data and outflow parameter data. It then forms a regression model between the outflow parameter data and inflow parameter data. This method is highly accurate, with average relative deviation of less than 10%. However, in scenarios where there is a sudden change in inflow parameters, such as shock loading, the model will not be able to accurately predict the outflow parameters due to the lack of historical data.

It would be desirable to overcome or ameliorate at least one of the abovedescribed problems with existing wastewater treatment processes, or at least to provide a useful alternative. Summary

The process disclosed herein is a hybrid model. It comprises a mechanistic model that runs prior to a statistical model. In this model, the mechanistic model generates an estimated output based on the measured input. The statistical model then uses the output of the mechanistic model and measured input for predicting the outflow parameter data.

The process can be used in the wastewater treatment industry. This hybrid model can be implemented for the prediction of the parameters of outflow of wastewater treatment processes for both domestic and industrial wastewater.

Disclosed is a system for estimating one or more outflow parameters, comprising: a mechanistic module comprising a mechanistic simulator; a machine learning module comprising a machine learning model; database module; and at least one processor, the database module storing instructions that, when executed by the at least one processor, cause the at least one processor to: receive wastewater inflow parameter data (inflow data) for a wastewater treatment plant; process the inflow data using the mechanistic simulator to produce a process estimation of a wastewater treatment process; and process at least part of each of the process estimation and inflow data using the machine learning model to produce an outflow estimation comprising at least one said outflow parameter.

The process estimation is an estimation or rough calculation of the wastewater treatment process. It contains effluent parameters. The at least one processor may be configured to process all the process estimation and inflow data. In this sense, the process estimation and inflow data are processed in their entirety. The set of inflow data received by the mechanistic simulator is thus the same as that received by the machine learning model.

The process estimation may comprise at least one estimated outflow parameter, and wherein the machine learning model is configured to process the process estimation and inflow data by correlating the inflow data with the process estimation.

The outflow estimation may change kinetics of at least one parameter of the process estimation.

The mechanistic model may be adapted to change the process estimation based on sudden changes in one or more inflow parameters of the inflow data. The sudden changes comprise a change in inflow wastewater volume. The sudden changes may comprise a change in capacity of the wastewater treatment plant.

The mechanistic simulator is configured to estimate at least one of mixed liquor volatile suspended solids, heterotroph concentration and effluent chemical oxygen demand.

The mechanistic simulator may be one of an activated sludge model and an anaerobic digestion model.

The one or more outflow parameters may comprise a plurality of output parameters and wherein a first set of said outflow parameters is produced by the mechanistic simulator and a second set of said outflow parameters is produced by the machine learning model. Also disclosed herein is a method for estimating one or more outflow parameters, comprising: receiving wastewater inflow parameter data (inflow data) for a wastewater treatment plant; processing the inflow data using a mechanistic simulator to produce a process estimation of a wastewater treatment process; and processing at least part of each of the process estimation and inflow data using a machine learning model to produce an outflow estimation comprising at least one said outflow parameter.

Processing at least part of each of the process estimation and inflow data may comprise processing all the process estimation and inflow data.

The process estimation may comprise at least one estimated outflow parameter, and wherein the machine learning model is configured to process the process estimation and inflow data by correlating the inflow data with the process estimation.

Processing at least part of each of the process estimation and inflow data using a machine learning model to produce an outflow estimation comprising at least one said outflow parameter, may comprise changing kinetics of at least one parameter of the process estimation.

The method may further comprise changing the process estimation based on sudden changes in one or more inflow parameters of the inflow data. Changing the process estimation based on sudden changes in one or more inflow parameters of the inflow data may comprise changing the process estimation based on a change in inflow wastewater volume.

Changing the process estimation based on sudden changes in one or more inflow parameters of the inflow data may comprise changing the process estimation based on a change in capacity of the wastewater treatment plant. The mechanistic simulator may be configured to estimate at least one of mixed liquor volatile suspended solids, heterotroph concentration and effluent chemical oxygen demand.

The mechanistic simulator may be one of an activated sludge model and an anaerobic digestion model.

The method may further comprise producing the one or more outflow parameters from a first set of said outflow parameters produced by the mechanistic simulator and a second set of said outflow parameters is produced by the machine learning model.

Advantageously, the presence of mechanistic model allows for mass balance related changes. In a shock change scenario, a change in the influent will not result in an abrupt change in the model prediction, but a gradual change in the model prediction. This information allows the operators to know how much time it will take for the full change to take effect, giving them time to make proper adjustment.

Advantageously, in a change in plant capacity scenario, embodiments of the invention can also calculate the performance better than the other hybrid models. Other hybrid models (e.g. Parallel hybrid model) work based on the assumption that the plants are operated within design capacity. The machine learning model in the parallel hybrid model only takes information from the influent, which then provides the correction factor to the mechanistic model. These embodiments, however, will capture the scenarios more accurately as the mechanistic model will provide additional information for the machine learning model. For example, in cases where the plant is overloaded, effluent ammonia will be detected in the effluent. This will be captured by the mechanistic model, which will provide information to the machine learning model. The machine learning model in this invention will have more information to use than a parallel hybrid model.

Brief description of the drawings

Embodiments of the present invention will now be described, by way of nonlimiting example, with reference to the drawings in which:

Figure 1 schematically illustrates a process in accordance with the present disclosure, and a system for performing that process; and

Figure 2 shows an example application of the process of Figure 1 for prediction of chemical oxygen demand (COD); and

Figure 3 shows an application of the process of Figure 1 for predicting effluent total nitrogen (TN) in a full-scale municipal wastewater treatment plant.

Detailed description

Disclosed are systems for predicting an effluent parameter associated with a wastewater treatment process including a mechanistic simulator configured to receive a first input dataset comprising a plurality of wastewater inflow parameters to provide an rough calculation of the wastewater treatment process; a predictor module configured to receive the rough calculation of the wastewater treatment process and the plurality of wastewater inflow parameters as a second input dataset to predict the effluent parameter.

The processes disclosed herein enable more accurate estimation of outflow parameters of the outflow of line a wastewater treatment plant. The process uses a mechanistic model and machine learning model. The mechanistic model is placed, in the sense of steps of a process, prior to the machine learning model. This makes the process more robust in the face of changes in plant influent

RECTIFIED SHEET (RULE 91) volume or parameters and the physical plant itself. The machine learning model correlates the inflow parameters an output of the mechanistic model to outflow parameter data. This significantly reduces labour and time required further COD fractionation. It also compensates for discrepancies that may not be captured by a conventional mechanistic model, making the process more accurate than mechanistic model.

The process significantly increases the accuracy of wastewater effluent prediction, without a significant increase in computing power. Additional 'modules' can be added after the model, increasing its functionality. This process is capable of optimizing certain processes in wastewater treatment systems, improving efficiency and potentially reducing operating cost.

Such a process 100 is shown in Figure 1. The process 100 arranges a mechanistic model or mechanistic simulator prior to a statistical (machine learning) model to predict the treatment performance of the wastewater treatment process. The mechanistic simulator provides a rough calculation of the treatment performance. The statistical model then calculates an accurate prediction of the treatment performance using the inflow parameters and the rough calculation as input.

The statistical model includes a machine learning model, which may be a regression-based model. The machine learning model can be trained and validated using the historical data from the treatment plant and data from the mechanistic model simulating the treatment plant. This is advantageous as the hybrid model uses the machine learning model to compensate the shortcomings of mechanistic model, increasing its accuracy. In addition, use of the machine learning model reduces the need for pilot-study and lab scale study of the plant. Conventionally, lab-scale study is required in order to understand the kinetics of the biomass as well as the biodegradability of the wastewater.

The process 100 broadly includes:

RECTIFIED SHEET (RULE 91 ) (Step 102) receiving wastewater inflow parameter data (inflow data - 114) for a wastewater treatment plant (not shown);

(Step 104) processing the inflow data using a mechanistic simulator 108 to produce a process estimation of a wastewater treatment process. The mechanistic simulator does a rough calculation of the wastewater treatment based on the inflow data; and

(Step 106) processing at least part of each of the process estimation produced at Step 104 and the inflow data received at Step 102 using a machine learning model or predictor module to produce an outflow estimation comprising at least one outflow parameter 112. The machine learning module 110 (hereinafter interchangeably referred to as the predictor module) receives the rough calculation from the mechanistic simulator and inflow parameters and uses both to predict an effluent outflow parameters. In general, the outflow estimation will include multiple outflow parameters.

The inflow parameter data 114 can include any parameters that are desired to be measured in the inflow data such as, but not limited to, COD, total organic carbon (TOC), solid content, ionic content, inorganic contaminants, and organic contaminants. The data 114 can be obtained through laboratory analysis, sensor data or any other means, and then received by the present process for outflow parameter estimation. Sensors, both physical and soft sensors, may be positioned around the wastewater treatment plant to facilitate the data collection process. Similarly, effluent or outflow parameters 112 are any parameters desired to be measured in the outflow including, but not limited to, one or more of a COD, nitrogenous content, phosphorous content and suspended solids.

The inflow parameter data 114 that, in the embodiment shown in Figure 1, forms part of a database module 116 that stores the data in non-transient memory for future use - historical data review or for retraining the in learning module 110. The database module 116 may include one or more computer

RECTIFIED SHEET (RULE 91 ) processors, computer networks or a cloud network or any other system or system component. The same may be said for the mechanistic module 104 and the machine learning module or predictor module 106.

Once the inflow parameter data 114 has been received, it is processed by the mechanistic module 108 in process step 104. The mechanistic module 104 runs a mechanistic simulator to simulate performance of the treatment system or plant and produce a process estimation 118 that is, or include, a prediction of outflow parameter data. Thus the mechanistic module 104 receives an updated or current influent parameter data 114 from database module 116.

The kinetic parameters of the mechanistic module 104 are first calibrated using either optimization of historical data or lab analysis. Calibration may be performed before receipt of the influent parameter data 114 by the mechanistic module 108, or calibration may be performed by the mechanistic module 108.

The mechanistic simulator may also require influent data that requires extensive calibration and lab analyses. For data that require extensive calibration and lab analyses, average data may be used. An example is the COD fraction of influent wastewater used in an activated sludge model. The COD fraction of the wastewater must be determined for the mechanistic simulator to make an accurate prediction. In this scenario, the average biodegradability can be used. The mechanistic module 108 can then generate the output data (process estimation 118), which is used as the input of the predictor module and predicted outflow parameter. In the scenario of simulating an activated sludge process, the input of the predictor module 110 that is generated by the mechanistic simulator can be mixed liquor volatile suspended solids, heterotroph concentration and effluent chemical oxygen demand, among others.

The mechanistic simulator may include, but is not limited to, one or more of activated sludge model or anaerobic digestion model. The mechanistic simulator

RECTIFIED SHEET (RULE 91 ) is used to simulate the treatment plant performance in the form of a process estimation. The process estimation produced by the mechanistic simulator can include treatment performance parameters such as effluent chemical oxygen demand (COD), effluent nitrogenous content, reactor tank's mixed liquor volatile suspended solids can be calculated. The simulation of the mechanistic model may be based on calibrated parameters or average parameters.

The mechanistic simulator may be a formulaic model or other form of model output based on current info data rather than requiring training based on historical data. As such, the mechanistic simulator run on the mechanistic module 108 contained the process estimation 118 rapidly, or in real-time, based on sudden changes in one or more inflow parameters of the inflow data. For example, a spike in influent or wastewater volume or changes in wastewater treatment plant capacity - e.g. decommissioning of an old wastewater treatment tank or the addition of a new wastewater treatment tank - will be immediately responded to by the mechanistic simulator. In contrast, a machine learning model would slowly adapt to the higher volume or changing in plant capacity over time, as data relating to the higher volume forms an increasingly large proportion of the training data from which the machine learning model learns.

The mechanistic simulator passes the process estimation 118 to the predictor module 110 for processing using the machine learning model. The machine learning model may be based on either supervised learning or unsupervised learning. The type of machine learning model may be selected based on the suitability of the machine learning model to the treatment plant or influent parameters. In addition, for different treatment plant, different database may be required as characteristics and constituents of the wastewater may vary for different treatment plant.

The input set to the machine learning model may include a list of wastewater inflow data such as chemical oxygen demand, total organic carbon, solids

RECTIFIED SHEET (RULE 91 ) content, ionic content, inorganic contaminant, and organic contaminant, and calculated output of the mechanistic simulator. The output generated by the machine learning model may be the effluent parameters of the treatment process.

The machine learning takes both the inflow parameter data from database module 116 and part or all of the process estimation 118 and finds one or more correlations therebetween. This is advantageous as the machine learning model is able to overcome the shortcoming of mechanistic simulator. Conventionally, the mechanistic simulator requires careful calibration, which requires lab-scale or pilot scale studies. The use of machine learning removes the calibration process, and is able to provide accurate simulation results. In this regard, the machine learning model may process each of the process estimation and info data to produce an outflow estimation but also to change kinetics or other characteristics of one or more parameters of the process estimation. This makes the present process 100 highly adaptable as the mechanistic simulator facilitate immediate response to sudden changes was the machine learning model refines the output based on historical wastewater treatment plant.

This significantly reduces the time required for accurate simulation and can result in accurate simulations. Effluent wastewater parameters such as chemical oxygen demand COD, nitrogenous content, suspended solids and phosphorous content can be predicted as the final output effluent wastewater parameters.

The predicted outflow parameters or outflow estimation 112 may be derived solely from the output of the predictor or machine learning module 110. Alternatively, the predicted outflow parameters 112 may be a combination of the output of the mechanistic model 108 and the machine learning or predictor module 110. For example, the predicted outflow parameters 112 may be a weighted sum or average of the process estimation 118 and the outflow estimation of the machine learning model. Alternatively, the predicted outflow parameters 112 may include one or more parameters from a first set of outflow

RECTIFIED SHEET (RULE 91 ) parameters calculated by the mechanistic simulator of the mechanistic module 108, and a second set of outflow parameters calculated by the machine learning model of the machine learning module 110. The first set and second set of outflow parameters may be non-overlapping, fully overlapping, or partially overlapping. Where there is any overlap, the one or more overlapping parameters may be combined into a single parameter using any desired method such as a weighted sum or average of the two values for each overlapping parameter.

In addition, the process estimation 118 may include only those parameters sought to be incorporated into the outflow estimation of learning model, or may include the entire output of. Similarly, the inflow parameters 114 sent to the predictor module 110 may include all inflow parameters or a subset of the influent parameters.

In view of the present disclosure, it will be understood that any of the above process steps may be performed on a single processor or on multiple processes, in a single computer system or multiple computer systems.

By merging the output of the mechanistic simulator with the inflow parameters for processing using machine learning model, a hybrid model is produced that can rapidly respond to changes in input parameters without losing the ability to refine kinetics of the outflow parameters based on historical data. The hybrid model is suitable for all biological wastewater treatment and can be used for prediction of one or more effluent parameters of biological wastewater treatment processes.

The results of the process 100 are shown in Figure 2. Figure 2 shows the measured effluent COD and the simulated effluent COD predicted by the hybrid model of process 100, as well as the simulated effluent COD produced using a commercial simulator. Clearly, the hybrid model of process 100 much more

RECTIFIED SHEET (RULE 91 ) closely tracks the actual output when compared with the model used by the commercial simulator.

Figure 3 shows effluent total nitrogen (TN) in a full-scale municipal wastewater treatment plant and the prediction made by the present methods. In both cases, prediction very closely tracks the actual measurements.

It will be appreciated that many further modifications and permutations of various aspects of the described embodiments are possible. Accordingly, the described aspects are intended to embrace all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims.

Throughout this specification and the claims which follow, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that that prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates.

RECTIFIED SHEET (RULE 91 )