Method and device for predicting a digester condition monitoring indicator

Abstract

A method (1) for predicting a monitoring indicator for the state of a digester adapted to produce digestate and biogas, containing methane, from an input, comprising: – A step of defining (10) a suitable mechanistic model to determine at least one value of a monitoring indicator for the state of said digester; – A step of determining (12) the tuning parameters of said mechanistic model describing the kinetics of at least one of the acidogenesis, acetogenesis, and methanogenesis stages, or their combination, as a function of control parameters of said digester (50) and parameters relating to the quality of the input; – A calibration step (13) of said mechanistic model; and – A step of implementing said mechanistic model to predict the evolution of said at least one value of a monitoring indicator for the state of said digester. Abstract figure: Figure 1

Description

Title of the invention: Method and device for predicting an indicator for monitoring the state of a digester

[0001] The invention relates to the field of digesters, also called biogas reactors or methanizers.

[0002] More particularly, the invention relates to a method and a device for predicting an indicator for monitoring the state of a digester.

[0003] The purpose of a digester is to produce biogas by anaerobic digestion.

[0004] Anaerobic digestion corresponds to a cascade of biochemical reactions allowing a fraction of organic matter to be converted into biogas, the remaining matter being called digestate.

[0005] Such a digestion process is inherently complex because:

[0006] - each stage of degradation is carried out by groups of microorganisms distinct, each having their own kinetics, optimal growth conditions, specific inhibitory compounds;

[0007] - these groups of microorganisms are in interaction with each other, with for example example of competition or syntrophy phenomena;

[0008] - the ecological balance and kinetics of these groups of microorganisms depend heavily on the operating conditions of the process, in particular its feeding method, its temperature or even the nature of the input used as feed.

[0009] This complexity makes it difficult to monitor an anaerobic digestion process. Thus, it is particularly complex for operators to predict the evolution of the state of a digester following a change in operating conditions.

[0010] Furthermore, the typical response time of a digester treating sludge and / or waste following, for example, a change of feed is of the order of several weeks.

[0011] Thus, it is essential to be able to anticipate the effects of a change in operation in order to limit the industrial risks.

[0012] To address this complexity, a variety of approaches exist in the state of the art.

[0013] The use of indicators based on physicochemical parameters for monitoring the process or characteristics of the substrate is known. This approach is purely empirical and often based on trial and error methods.

[0014] For example, we know from the publication C, Veluchamy & Gilroyed, Brandon & Kalamdhad, Ajay. (2019). Process performance and biogas production optimizing of mesophilic plugflow anaerobic digestion ofcom silage. Fuel. 253. 1097-1103. 10.1016 / j.fuel.2019.05.104. the calculation of the FOS / TAC ratio (ratio between volatile acidity and total alkalinity) to estimate whether a digester is overloaded. Thus we find decision tables such as this one: FOS / TAC Ratio Situation Action to be taken to remedy the situation >0.6 Very excessive biomass input Stop adding biomass 0.5-0.6 Excessive biomass input Add less biomass 0.4-0.5 Reactor is heavily loaded Monitor reactor more closely 0.3-0.4 Biogas production is maximal Keep biomass input constant 0.2-0.3 Biomass input is too low Slowly increase biomass input <0.2 Biomass input is much too low Rapidly increase biomass input

[0015] We can also cite the monitoring of the percentage of CH4 in the biogas, the concentration of total ammoniacal nitrogen or even the carbon / nitrogen ratio, more commonly called the C / N ratio.

[0016] These indicators have the advantage of being simple to put into practice, but have a low predictive capacity and are only used in practice for the purposes of digester stabilization rather than optimization.

[0017] We also know the implementation of so-called "black box" models based on the monitoring parameters of the digestion process. A "black box" model is known as a type of model that does not take into account the phenomena internal to the system studied, here the digester.

[0018] These models make it possible to predict a parameter of interest, such as methane yield, from the operating conditions of the digester and the characteristics of the substrate.

[0019] These are generally models of relatively simple structure, such as linear models, for example multilinear regression models, partial least squares regression, known as PLS, or power (y=a.xb) models, although any model structure can be used, including artificial neural networks, SVR support vector regressions, etc.

[0020] Such models are generally very specialized and only valid for restricted operating and power supply conditions.

[0021] The use of "black box" models based on bioindicators is also known. This approach uses data generated by molecular biology tools such as quantitative PCR, called qPCR, or DNA sequencing.

[0022] The principle is identical to the “black box” models based on process monitoring parameters, but is based on the predictive power of bio-indicators, for example to predict the concentration of a bacterial species.

[0023] Models are then trained to link these bioindicators with behaviors observed in the digesters.

[0024] For example, a model for predicting the degree of inhibition to NH3 or phenol from sequencing data was proposed in the publication Poirier, S., Déjean, S., Midoux, C., Lê Cao, K.-A., Chapleur, O., 2020. Integrating independent microbial studies to build predictive models of anaerobic digestion inhibition by ammonia and phenol. Bioresource Technology 316, 123952. These models are, however, little developed, due to the complexity of processing sequencing data, as well as the high cost of sequencing compared to conventional physicochemical parameters.

[0025] We also know the use of mechanistic models adjusted on real data.

[0026] The phenomena governing the different stages of anaerobic digestion are for most well known and can be described mathematically in the form of dynamic models. One of the most widely used models in the field is ADM1 and its variants, presented in its original form in the publication Batstone, DJ, Keller, J., Angelidaki, I., Kalyuzhnyi, SV, Pavlostathis, SG, Rozzi, A., Sanders, WTM, Siegrist, H., Vavilin , VA, 2002. The IWA Anaerobic Digestion Model No 1 (ADM1 ). Water Science and Technology 45, 65-73. This model has the advantage of explicitly describing each stage of anaerobic digestion and of integrating a large number of factors having an influence on the different kinetics.

[0027] However, this type of model generally includes a very high number of parameters, such as biological kinetics parameters, inhibition constants, yields, which must be adjusted using real data. When used for predictive purposes, these models are used as follows:

[0028] - Model parameters are adjusted from existing data, e.g. monitoring data from the last few months of a digester, minimizing the gap between model outputs, such as biogas flow rate, and actual measurements, and

[0029] - Assuming these parameters are constant over time, the model can then be used to predict the future behavior of the digester, for example by changing the mode or composition of the feed.

[0030] However, the adjustment of these models is regularly subject to overfitting due to the large number of parameters that can compensate for each other. others. Moreover, the predictions of this type of model are generally only valid for the digester used for the adjustment.

[0031] Finally, we know the implementation of hybrid models adjusted on real data.

[0032] In order to limit the risks of overlearning and increase the predictive power of mechanistic models, the latter can be coupled with empirical models of the “black box” type.

[0033] Today, these approaches are mainly implemented through software sensors, in English soft-sensors, which consist of estimating an unmeasured parameter from the signals acquired online on a process; as explained in the publication Haimi, H., Mulas, M., Corona, F., Vahala, R., 2013. Data-derived soft-sensors for biological wastewater treatment plants: An overview. Environmental Modelling & Software 47,88-107.

[0034] Another approach is to use data external to the process to inject them into the model. For example, the publication Charnier, C., Latrille, E., Jimenez., J., Torrijos, M., Sousbie, P., Miroux, J., Steyer, J.-P., 2017. Fast ADM1 implementation for the optimization offeeding strategy using near infrared spectroscopy. Water Research 122, 27 - 35 discloses a method in which a partial least squares approach model, called PLS, was trained, which can predict hydrolysis kinetics of a substrate from its near infrared spectrum.

[0035] This coupling therefore makes it possible to reduce the number of parameters to be adjusted on real data as well as to inject knowledge from past experiences. However, this type of mixed approach requires a large number of external data to be integrated in order to optimize its operation.

[0036] Also, there is a need for a solution for predicting the evolution of the state of a digester that solves the problems of the known methods of the prior art.

[0037] To this end, a method is proposed for predicting an indicator for monitoring the state of a digester suitable for producing digestate and biogas, containing methane, from an input, comprising:

[0038] - A step of defining a suitable mechanistic model to determine the at least one value of an indicator for monitoring the state of said digester as a function of control parameters of said digester and parameters relating to the quality of the input feeding said digester; said mechanistic model being intended to be calibrated by a plurality of adjustment parameters describing the kinetics of the stages of hydrolysis, acidogenesis, acetogenesis and methanogenesis, independently or in combination;

[0039] - A step of acquiring values ​​of the control parameters of said digester and of quality parameter values ​​of the input feeding said digester;

[0040] - A step of determining the adjustment parameters of said mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination as a function of the control parameters of said digester and the parameters relating to the quality of the input;

[0041] - A step of calibrating said mechanistic model as a function of said parameters of settings obtained by the implementation of said machine learning module; and

[0042] - A step of implementing said mechanistic model to predict the evolution of said at least one value of an indicator for monitoring the state of said digester.

[0043] Thus, the evolution of the state of a digester can be predicted in a relatively robust manner by implementing in a relatively optimal manner a mechanistic module simulating the internal processes of the digester.

[0044] Indeed, mechanistic models allow the injection of knowledge based on explicit hypotheses, such as compliance with the material balance, and ensure a result that respects the laws of physics.

[0045] Advantageously, said determination step is implemented by a machine learning module.

[0046] In particular, said machine learning module comprises a neural network or a decision tree or forest of decision trees module or a partial least squares approach module or a support vector regression module.

[0047] Advantageously and in a non-limiting manner, the adjustment parameters of the mechanistic model are constrained within a range of predetermined values ​​or constrained according to a predetermined distribution of the adjustment parameters. Thus, it can be ensured that the mechanistic model is not trained with aberrant or obviously erroneous adjustment parameter values.

[0048] Advantageously and in a non-limiting manner, said indicator for monitoring the state of said digester comprises a flow rate of biogas produced at the outlet of the digester. Thus, a relatively relevant indicator of the state of the digester and in particular of its performance can be obtained.

[0049] Advantageously and in a non-limiting manner, said automatic learning module comprises a neural network or a decision tree or forest of decision trees module or a partial least squares approach module or a support vector regression module. Thus, a robust and efficient module is implemented to determine the adjustment parameters of the mechanistic model.

[0050] Advantageously and in a non-limiting manner, said parameters relating to the input comprise a methanogenic potential parameter and / or parameters of splitting. These parameters are in fact relevant indicators of the quality of the input.

[0051] Advantageously and in a non-limiting manner, said control parameters of the digester comprise a temperature of said digester and / or a hydraulic residence time and / or an organic load and / or a feeding method of said digester. These parameters in fact constitute parameters which can be directly adjusted by the operators and / or designers of the digester.

[0052] The invention also relates to a method for training said automatic learning module for a prediction method as described previously, characterized in that it comprises:

[0053] - A step of acquiring a plurality of experimental data sets, each set of experimental data comprising values ​​of control parameters of a digester, values ​​of parameters relating to the quality of a type of input and values ​​of said monitoring indicator of said digester as a function of these parameters;

[0054] - For each experimental data set, a calibration step of a mechanistic model, in which the values ​​or distributions of the adjustment parameters of said mechanistic model are determined as a function of said experimental data;

[0055] - A step of training said automatic learning module in based on said plurality of experimental data sets, said machine learning module being trained to determine said adjustment parameters of said mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination based on the values ​​of control parameters of the digester and the values ​​of parameters relating to the quality of the input.

[0056] This thus makes it possible to efficiently train a machine learning module to determine adjustment parameters of a mechanistic model from experimental data previously acquired from one or more existing digesters.

[0057] The invention also relates to a method for regulating the control parameters of a digester comprising:

[0058] - The acquisition of a prediction obtained by the implementation of a method of prediction as previously described; and

[0059] - Regulation of the digester control parameters as well as the feed by inputting said digester based on said acquired prediction.

[0060] The invention also relates to a device for training a machine learning module to predict the evolution of the state of a digester adapted to produce digestate and biogas from an input comprising:

[0061] - Means for acquiring a plurality of sets of experimental data, each set of experimental data comprising values ​​of parameters relating to a digester, values ​​of parameters relating to a type of input and a value of said indicator for monitoring the state of said digester as a function of these parameters;

[0062] - For each set of experimental data, means of calibrating a mechanistic model adapted to determine at least one value of an indicator for monitoring the state of a digester as a function of parameters relating to said digester and parameters relating to the input, said mechanistic model being calibrated by determining the values ​​of a plurality of adjustment parameters describing the kinetics of the stages of hydrolysis, acidogenesis, acetogenesis and methanogenesis, independently or in combination, as a function of said experimental data; and

[0063] - Means for training at least one automatic learning module based on said plurality of experimental data sets, said machine learning module being trained to determine said adjustment parameters of said mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination based on the values ​​of parameters relating to said digester and the values ​​of parameters relating to the input.

[0064] The invention also relates to a device for implementing the prediction of a machine learning module driven by a device as described previously, to predict the evolution of the state of a digester adapted to produce digestate and biogas from an input comprising:

[0065] - means for acquiring values ​​of parameters relating to said digester for which the status monitoring indicator value is to be predicted and the input parameter values ​​provided to said digester,

[0066] - means for implementing said automatic learning module of so as to obtain adjustment parameters of a mechanistic model as a function of said parameters acquired in the previous step; said mechanistic model being the same as that implemented by the prediction device described previously;

[0067] - means for calibrating said mechanistic model as a function of said parameters of settings obtained by the implementation of said machine learning module; and

[0068] - means of implementing said mechanistic model to predict the evolution of at least one indicator for monitoring the state of said digester.

[0069] The invention also relates to a device for controlling a digester adapted to regulate parameters of said digester and / or regulation parameters of the mode feed of the digester according to the prediction obtained by said prediction method described above.

[0070] Other features and advantages of the invention will emerge from reading the description given below of a particular embodiment of the invention, given for information purposes, but not as a limitation, with reference to the appended drawings in which:

[0071] [Fig-1] is a flowchart of the prediction method according to the invention;

[0072] [Fig.2] is a flowchart of a training method according to the invention;

[0073] [Fig.3] is a schematic view of a device for implementing a method of prediction according to the invention;

[0074] [Fig.4] is a schematic view of a device for implementing a method training according to the invention; and

[0075] [Fig.5] is a schematic view of a method for regulating the parameters of control of a digester.

[0076] The prediction method 1 according to the invention aims to predict the evolution of the state of a digester, fed by an input such as sludge from a treatment plant and producing digestate and biogas at the output. The prediction is thus implemented for a particular digester 50, here called the digester 50 studied.

[0077] In fact, each digester, due to the anaerobic reactions involved, presents different performances and a different evolution of its state, depending on the input which feeds it and depending on its control parameters.

[0078] The evolution of the state of a digester 50 also presents a strong inertia and a slow evolution over time, which causes particular difficulties in anticipating the degradation of its performance in order to modify its input supply.

[0079] The method comprises a first step of defining a mechanistic model making it possible to predict the evolution of an indicator for monitoring the state of the digester.

[0080] The definition of the mechanistic model can be a choice made by the user implementing the prediction method 1. However, as will be seen later, this choice must in reality have been made before the training 2 of the automatic learning module, because the training is dependent on the chosen mechanistic model.

[0081] In this embodiment, the flow rate of biogas produced at the outlet of the digester 50 is the indicator for monitoring the state that we seek to predict, because it constitutes an important indicator in terms of profitability of the digester. However, the invention is not limited to this particular type of indicator, which may also correspond, in a non-limiting manner, to the concentration of volatile fatty acids in the digestate, or even the concentration of NH3.

[0082] The scientific literature includes numerous mechanistic models adapted to anaerobic digesters.

[0083] Among the well-known models of the prior art, the Anaerobic Digestion Model No. 1 (ADM1) and its variants are known. Such models make it possible in particular to simulate the anaerobic reactions of a digester 50 to predict the state of the digester, its performance or even its stability.

[0084] The invention is not limited to a particular mechanistic model. Indeed, as will be explained later, the invention will train a machine learning module to determine the optimal adjustment parameters of the mechanistic model in order to be able to best simulate the internal reactions of the digester 50 studied.

[0085] Also, any mechanistic model adjustable by a plurality of predefined parameters can be implemented.

[0086] However, in order to limit the amount of calculations required for automatic learning of the adjustment parameters, it is advantageous to select a mechanistic model comprising a reduced number of adjustment parameters; this is then referred to as a reduced mechanistic model.

[0087] For this purpose, methods for reducing mechanistic models are known.

[0088] Examples include the following publications:

[0089] - B. Chachuat, N. Roche, and MA Lad . Reduction of the ASM1 model for the Optimal control of small activated sludge treatment plants. Journal of Water Sciences, 16(1):526, 2003; or

[0090] - Sonia Hassam. Reduction of biotechnological process models: Applications to ADM1. Automation / Robotics. University of Tlemcen, 2015. French. ( tel -01249831 ).

[0091] However, the person skilled in the art will again be free to select the mechanistic model of his choice and, where appropriate, the associated reduction method.

[0092] Once the mechanistic model has been defined, and the adjustment parameters are known, a step 11 is implemented for acquiring values ​​of control parameters of the digester 50 for which the monitoring values ​​are to be predicted and values ​​of quality parameters of the input feeding said digester.

[0093] In other words, we acquire the measurements which will be provided to the mechanistic model to determine the evolution over time of the digester's performance.

[0094] For this purpose, the values ​​provided to the mechanistic model are generally time series, respectively of control of the digester 50 and quality of the input.

[0095] As non-limiting examples, we can cite as quality parameters of the input its methanogenic potential, which is a relatively important criterion, its chemical oxygen demand.

[0096] The quality of the input can also include its fractionation parameters into sugars, proteins, lipids, metallic elements, elemental composition, etc.

[0097] Regarding the parameters relating to the digester 50, the control parameters of the digester 50 may include, by way of non-limiting examples, the temperature of the digester, the hydraulic residence time, the organic load, the mode of feeding said digester, etc.

[0098] Then, before providing these input parameters to the mechanistic model, it is necessary to calibrate the adjustment parameters of the mechanistic model.

[0099] To do this, a step 12 is implemented for determining the adjustment parameters of said mechanistic model.

[0100] The adjustment parameters of a mechanistic model for an anaerobic digester will be used to calibrate the representation of the kinetics of the stages of anaerobic decomposition, namely: the stages of hydrolysis, acidogenesis, acetogenesis and methanogenesis.

[0101] Hydrolysis refers to the step during which high molecular weight organic molecules such as polysaccharides, proteins and lipids are converted into monomers (e.g. glucose, fatty acids, amino acids). This step is commonly described by first-order kinetics associated with a hydrolysis constant. In particular, one can refer for the description of this hydrolysis kinetics to the publication Vavilin, VA, Fernandez, B., Palatsi, J., Flotats, X., 2008. Hydrolysis kinetics in anaerobic degradation of particulate organic material: An overview. Waste Management 28, 939-951. https: / / doi.Org / 10.1016 / j.wasman.2007.03.028

[0102] Acidogenesis corresponds to the step during which the monomers resulting from the hydrolysis step are transformed into organic acids, alcohols, H2 and CO2. The kinetics of this step can for example be described according to the empirical Monod equation, characterized by a maximum growth rate and an affinity constant. The Monod equation is notably described in the publication Liu, Y., 2007. OverView of some theoretical approaches for derivation of the Monod equation. Appl Microbiol Biotechnol 73, 1241 - 1250. https: / / doi.org / 10.1007 / s00253-006-0717-7.

[0103] The acetogenesis phase allows the transformation of the different organic compounds from the two previous phases into acetate, H2 and CO2. The kinetics of this phase can also be described by Monod kinetics corrected for the accumulation of reaction products such as H2. The Monod kinetics thus corrected is notably described in the publication Patôn, M., Rodriguez, J-, 2019. Integration of bioenergetics in the ADM1 and its impact on model predictions. Water Sci Technol 80, 339-346. https: / / doi.org / 10.2166 / wst.2019.279.

[0104] Finally, the methanogenesis step corresponds to the conversion of H2, CO2 and / or acetate into CH4. The kinetics of this step can be described by Monod kinetics corrected for inhibition linked to toxic compounds such as nitrogen ammoniacal. For a description of the kinetics of this stage of methanogenesis, we can refer in particular to the publication Capson-Tojo, G., Moscoviz, R-, Astals, S., Robles, A., Steyer, J.-P., 2020. Unraveling the literature chaos around free ammonia inhibition in anaerobic digestion. Renewable and Sustainable Energy Review s 117, 109487. https: / / doi.org / 10.10167j.rser.2019.109487.

[0105] According to the invention, during the determination step, the adjustment parameters of the mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination are determined.

[0106] Advantageously, according to a particular implementation of the invention, all the adjustment parameters of the mechanistic model can be determined.

[0107] In this embodiment, the determination step 12 is implemented by implementing a machine learning module trained to determine the adjustment parameters as a function of the control parameters of the digester 50 studied and the parameters relating to the quality of the input.

[0108] This automatic learning module thus receives as input control parameters of the digester 50 studied and parameters relating to the quality of the input and determines as output adjustment parameters of the digester.

[0109] The machine learning module comprises in particular a plurality of artificial neural networks trained to each detect a setting parameter of the mechanistic model.

[0110] An alternative approach may implement an artificial neural network trained to output a single vector of adjustment parameters.

[0111] The invention is not limited to the implementation of neural networks; other machine learning methods are equally relevant for this type of learning. These include statistical regression methods such as support vector machines, SVMs, decision trees, decision tree forests, partial least squares regression, or any other machine learning method, whether supervised or not.

[0112] At the output of the implemented automatic learning module, the adjustment parameters of the mechanistic model are then recovered. Consequently, a calibration step 13 of the mechanistic model is implemented with these determined parameters.

[0113] During this calibration step 13, however, we want to ensure that we do not adjust the mechanistic model with aberrant parameter values. We also constrain the search for parameter values ​​within a predetermined distribution.

[0114] Alternatively, one could also plan to constrain the parameters between predetermined limits, this approach being able to be seen as a special case. of the search for parameters within a distribution, that corresponding to a distribution according to a uniform law included between these predetermined limits.

[0115] This then makes it possible to implement 14 the mechanistic model thus calibrated, by providing it as input with the control parameters of the digester 50 studied and parameters relating to the quality of the input.

[0116] In order to be able to implement this prediction method, however, it is necessary to have previously trained the machine learning module implemented.

[0117] To this end, a method 2 for training the automatic learning module is implemented.

[0118] This method comprises a first step 21 of acquiring a plurality of sets of experimental data.

[0119] Each set of experimental data comprises time series of values ​​of control parameters of a digester, of values ​​of parameters relating to the quality of a type of input feeding this digester 50 and time series of values ​​of the monitoring indicator of the digester, here its methane flow rate as a function of these parameters.

[0120] These different sets of experimental data are preferably acquired for a plurality of different digesters, and over time periods extending over several months. Thus, it is possible to obtain experimental data making it possible to know the effect over time on the digester monitoring indicator, for example here the methane flow rate, as a function of the input, its quality and as a function of control parameters of this particular digester 50.

[0121] The greater the quantity and variety of these experimental data sets, the greater the quality of machine learning can be.

[0122] Starting from these experimental data, we proceed to a calibration step 22 of a mechanistic model, which must of course be the same mechanistic model as the one that we will define during the definition step 10 of the prediction method 1.

[0123] We will therefore carry out this calibration by searching for the adjustment parameters allowing the input data to correspond with the evolution observed experimentally on the digester monitoring indicator.

[0124] Thus, we will take advantage of the knowledge of this experimental data to find an optimized parameterization of the mechanistic model as a function of the control parameters of the digester 50 and the quality parameters of the input.

[0125] The adjustment parameters obtained during this calibration step 21 constitute, with the set of experimental data associated with it, a set of training data for the automatic learning module.

[0126] We will then proceed to a training step 23 of the learning module as a function of the training data, constituted by the pairs of experimental data sets / adjustment parameters of the determined mechanistic model.

[0127] In this way the learning model will be able to evaluate, as a function of control parameters of said digester 50 and parameters relating to the quality of the input feeding said digester, the adjustment parameters of the chosen mechanistic model in an optimized manner.

[0128] This prediction method 1 can be implemented within the framework of more general methods, such as a method 53 for regulating the control parameters of a digester 50 shown [Fig.5].

[0129] Such a regulation method 53 comprises the acquisition of a prediction obtained by implementing the prediction method 1.

[0130] From this prediction, it is possible to regulate 51 the supply of input to the digester 50 as well as to vary the control parameters of the digester 50 to improve the performance of this digester.

[0131] Prediction method 1 can also be implemented to simulate the creation of a new digester, thereby estimating the performance of a project with a view to its design.

[0132] The device 40 for implementing the prediction method 1 is preferably a computer, comprising a processor 42 for carrying out the calculations, a random access memory 41, a data storage medium 43, input-output communication means 46, these means being adapted to store in memory, either in random access memory 41 or on the storage medium 43, a mechanistic model 48 and an automatic learning module 45.

[0133] The input-output communication means 46 can be of any known type, Ethernet port, COM port, USB port, wired or wireless communication device.

[0134] The control device of a digester 50 is generally, but not necessarily, of the same type as the implementation device 40 of the prediction method 1.

[0135] The training device 30 of the training method 2 is generally, but not necessarily, a device distinct from the implementation device 40 of the prediction method 1.

[0136] This training device 30 also comprises a computer, comprising a processor 42 for carrying out the calculations, a random access memory 41, a data storage medium 43, input-output communication means 46, these means being adapted to store in memory, either in random access memory 41 or on the storage medium 43, a mechanistic model 48 and an automatic learning module 45.

[0137] In the same way, the input-output communication means 46 can be of any known type, Ethernet port, COM port, USB port, wired or wireless communication device.

[0138] Furthermore, the training device 30 can implement the training steps of the automatic learning module with parallel computing processors, such as GPUs implemented in a general computing context, known in English by the acronym GPGPU, for General-Purpose computing on Graphics Processing Units.

[0139] According to an alternative implementation of the invention, the adjustment of the mechanistic models 38, 48 can be carried out using Bayesian approaches in order to transform the deterministic mechanistic model into a stochastic model.

Claims

1.

2. Claims Computer-implemented method (1) for predicting the evolution of an indicator for monitoring the state of a digester (50) adapted to produce digestate and biogas, containing methane, from an input; said method comprising: • A step of defining (10) a mechanistic model suitable for determining at least one value of an indicator for monitoring the state of said digester (50) as a function of control parameters of said digester (50) and parameters relating to the quality of the input feeding said digester (50); said mechanistic model being intended to be calibrated by a plurality of adjustment parameters describing the kinetics of the stages of hydrolysis, acidogenesis, acetogenesis and methanogenesis, independently or in combination; • A step of acquiring (11) values ​​of the control parameters of said digester (50) and values ​​of quality parameters of the input feeding said digester (50); • A step of determining (12) the adjustment parameters of said mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination exclusively as a function of the control parameters of said digester (50) and the parameters relating to the quality of the input; • A calibration step (13) of said mechanistic model as a function of said adjustment parameters obtained by the implementation of said automatic learning module; and • A step of implementing (14) said mechanistic model to predict the evolution of said at least one value of an indicator for monitoring the state of said digester (50); said determining step (12) being implemented by a machine learning module trained from experimental data from a plurality of different digesters. Prediction method (1) according to claim 1, characterized in that said machine learning module comprises a neural network or a decision tree or forest of trees module decision-making or a partial least squares approach module or a support vector regression module.

3. Prediction method (1) according to any one of claims 1 or 2, characterized in that during the calibration step (13) the adjustment parameters of the mechanistic model are constrained within a range of predetermined values ​​or according to a predetermined distribution of the adjustment parameters.

4. Prediction method (1) according to any one of claims 1 to 3, characterized in that said indicator for monitoring the state of a digester (50) comprises a flow rate of biogas comprising methane produced at the outlet of the digester.

5. Prediction method (1) according to any one of claims 1 to 4, characterized in that said parameters relating to the input comprise a methanogenic potential parameter and / or fractionation parameters.

6. Prediction method (1) according to any one of claims 1 to 5, characterized in that said control parameters of the digester (50) comprise a temperature of said digester (50) and / or a hydraulic residence time and / or an organic load and / or a feeding mode of said digester.

7. A computer-implemented training method (2) of a machine learning module for a prediction method according to any one of claims 1 to 6, characterized in that it comprises: • A step of acquiring (21) a plurality of experimental data sets from a plurality of different digesters, each experimental data set being associated with a digester of the plurality of different digesters, each experimental data set comprising values ​​of control parameters of a digester, values ​​of parameters relating to the quality of the inputs of said digester and values ​​of said monitoring indicator of said associated digester as a function of these parameters; • For each experimental data set, a step of calibrating (22) a mechanistic model (38) adapted to determine at least one value of an indicator monitoring the state of the digester associated with the experimental data set as a function of parameters relating to said associated digester and parameters relating to the input, said mechanistic model (38) being calibrated by determining the values ​​of a plurality of adjustment parameters describing the kinetics of the hydrolysis, acidogenesis, acetogenesis and methanogenesis steps, independently or in combination, as a function of said experimental data; • A step of training (23) said machine learning module as a function of said plurality of experimental data sets, said machine learning module being trained to determine for each experimental data set said adjustment parameters of said mechanistic model describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination exclusively as a function of the control parameters of said associated digester and the parameters relating to the quality of the input.

8. A computer-implemented method (53) for regulating control parameters of a digester (50) comprising: - The acquisition of a prediction obtained by implementing a prediction method (1) according to any one of claims 1 to 6; and - The regulation (51) of the control parameters of the digester (50) as well as the supply of input to said digester (50) according to said acquired prediction.

9. Training device (30) of a machine learning module for predicting an indicator for monitoring the state of a digester (50) adapted to produce digestate and biogas, containing methane, from an input comprising: • Means for acquiring (36) a plurality of sets of experimental data from a plurality of different digesters, each set of experimental data being associated with a digester of the plurality of different digesters; each set of data

10. experimental including values ​​of parameters relating to a digester, values ​​of parameters relating to a type of input and a value of said indicator for monitoring the state of the associated digester as a function of these parameters; • For each set of experimental data, means (32) for calibrating a mechanistic model (38) adapted to determine at least one value of an indicator for monitoring the state of the digester associated with the set of experimental data as a function of parameters relating to said associated digester and parameters relating to the input, said mechanistic model (38) being calibrated by determining the values ​​of a plurality of adjustment parameters describing the kinetics of the hydrolysis, acidogenesis, acetogenesis and methanogenesis steps, independently or in combination, as a function of said experimental data; and • Means for training (32) at least one machine learning module (45) as a function of said plurality of experimental data sets, said machine learning module (45) being trained to determine, for each experimental data set, said adjustment parameters of said mechanistic model (38) describing the kinetics of at least one of the stages of acidogenesis, acetogenesis and methanogenesis or their combination exclusively as a function of the control parameters of said associated digester and of the parameters relating to the quality of the input. Device for implementing (40) the prediction of a machine learning module (45) driven by a training device (30) according to claim 9, to predict the evolution of the state of said digester (50) adapted to produce digestate and biogas from an input comprising: • means for acquiring (46) parameter values ​​relating to said digester (50) for which the status monitoring indicator value is to be predicted and parameter values ​​of the input supplied to said digester, means for implementing (42) said automatic learning module (45) so as to obtain adjustment parameters of a mechanistic model (48) as a function of said parameters acquired in the previous step; said mechanistic model (48) corresponding to the mechanistic model implemented by the training device (30) according to claim 9; means (42) for calibrating said mechanistic model as a function of said adjustment parameters obtained by the implementation of said automatic learning module; and means (42) for implementing said mechanistic model to predict the evolution of at least one indicator for monitoring the state of said digester.

Images