A new methodology for early BMP assessment using a mathematical model
Sabrina Gu ´erin 2 St ´ephane Mottelet 1 Sam Azimi 2 Jean Bernier 2 Laura Andr ´e 3 Thierry Ribeiro 3 Andr ´e Pauss 1 Vincent Rocher 2
1
Universit ´e de Technologie de Compi `egne, FRANCE
2
SIAAP Direction D ´eveloppement et Prospective, Colombes, FRANCE
3
UniLaSalle Beauvais, FRANCE
15 th Anaerobic Digestion 2017 Conference
MOCOPEE Program (Modeling, Control and Optimization of Wastewater Treatment Processes, www.mocopee.com)
The Mocopee research program aims to build the metrological and mathematical tools (signal processing, treatment processes modeling, regulation) required to improve the control and the optimization of water and sludge treatment plants.
R&D actions on the AD process :
I
validate at industrial scale sewage sludge BMP estimation methods
I
build predictive models of AD process
The different substrates considered in the study
Primary, biologic (nitrification, denitrification), floated, mixed and thickened sludge, from different plants of SIAAP (Seine centre, Seine aval, Seine Gr ´esillons)
Inoculum sampled at the output of the digester
preliminary work :
S. Gu ´erin et al. (2016), Cartographie des boues de STEP et r ´eduction du temps de mesure du potentiel m ´ethanog `ene :
couplage exp ´erimentation en r ´eacteur / mod ´elisation
, L’eau, l’industrie, les
nuisances, n
o397
Experimental device
500 ml reactors,I/S ratio=3 CO 2 trapping
Mean flow measurement by ≈ 10ml throttles
AMPTS
Full compliance with experts recommendations :
C. Holliger et al. (43 auteurs) (2016), Towards a standardization of biomethane potential tests, Water
Science &Technology
Experimental study
36 batchs in triplicates MS, MV, DCO, DBO measurement
BMP obtained after 20 days No evident correlation between BMP and a priori measurements !
Could a model of AD digestion help to make an early assessment of BMP ?
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
0 2 4 6 8 10
t (day) 0
1 2 3
CH4 flow (g COD/L/day)
Modified AM2 model
O. Bernard, Z. Hadj-Sadok, D. Dochain, A. Genovesi, J. P. Steyer (2001), Dynamical model development and parameter identification for an anaerobic wastewater treatment process, Biotechnology and
bioengineering
R. Fekih Salem, N. Abdellatif, T. Sari, and J. Harmand (2012), On a three step model of anaerobic digestion including the hydrolysis of particulate matter, MATHMOD 2012
A. Donoso-Bravo, S. P ´erez-Elvira and F. Fdz-Polanco (2014), Simplified mechanistic model for the two-stage anaerobic degradation of sewage sludge, Environmental Technology
S 0 −−−−→ r
0S 1 , (Hydrolysis)
S 1 −−−−→ r
1Y X
1X 1 + (1 − Y X
1)S 2 + k 4 CO 2 , (Acidification) S 2 −−−−→ r
2Y X
2X 2 + (1 − Y X
2)CH 4 + k 5 CO 2 , (Methanogenesis) S 0 : insoluble organic molecules, S 1 : simple compounds (fatty acids, peptides, amino acids, . . .), S 2 : volatile fatty acids
Warning : we use the
batch
version of this model !
Modified AM2 model
Differential equations system
Perfectly mixed batch reactor, states of the sytem : S 0 , S 1 , S 2 , X 1 , X 2 S 0 0 = −r 0 , S 1 0 = r 0 − r 1 , S 2 0 = (1 − Y X
1)r 1 − r 2 , X 1 0 = Y X
1r 1 , X 2 0 = Y X
2r 2 , CH 4 0 = (1 − Y X
2)r 2 initial conditions : S 0 (0) = S 0 0 , S 1 (0) = S 1 0 , X 1 (0) = X 1 0 , X 2 (0) = X 2 0 . reaction rates :
r 0 = µ 0 S 0 , r 1 = µ max 1 S 1 X 1
S 1 + K S
1, r 2 = µ max 2 S 2 X 2
S 2 + K S
2+ S 2 2 /K I · Parameters θ = (Y X
1, Y X
2, µ 0 , µ max 1 , µ max 2 , K S
1, K S
2, K I
| {z }
θ
c: kinetic parameters
, X 1 0 , X 2 0 , S 0 0 , S 1 0 , S 2 0
| {z }
θ
b: batch parameters
)
Modified AM2 model
Identifiability
Knowing CH 4 (t), ∀t > 0 can we uniquely identify parameters ? I O. Bernard et al. (2001), R. Fekih Salem et al. (2012) model : no
I A. Donoso et al. (2014) : a priori no , but certain algebraic expressions are identifiable :
CH4(∞) = (1 − Y X
2)
(1 − Y X
1)(S 0 0 + S 1 0 ) + S 2 0
Other interesting expressions (relative proportions of substrates) : S 0 0
P 2
i=0 S i 0 , S 1 0 P 2
i=0 S 0 i , S 2 0 P 2
i=0 S i 0
Modified AM2 model
Practical identification
Goals :
1
obtain a mathematical model allowing to reproduce the methane rate of our 108 experiences, without necessarily uniquely describe state variables
(X 1 , X 2 , S 1 , S 2 , S 0 ).
2
being able to use this model to predict the BMP from new data measured after 4 days Pitfalls to bypass :
I No identifiability of parameters = ⇒ numerical problems !
I Important mass of data (108 experiences to assimilate)
Identification of parameters
Available measurements, simulations
For each batch #i we have
I (t k i ) k=1...m
ithe times of throttle switchs
I (D k i ) k=2...m
ithe mean CH 4 flow rate measured at t = t k i , k = 1 . . . m i For θ = (θ c , θ b ) we can simulate the mean flow of CH 4 :
d k i (θ c , θ b ) =
CH 4 (t k i ) − CH 4 (t k−1 i )
/(t k i − t k−1 i ),
and the function
J i (θ c , θ b ) =
m
iX
k=2
(t k i − t k−1 i )(D k i − d k i (θ c , θ b )) 2
evaluates the misfit between measurements of batch #i and the simulation with
parameters θ = (θ c , θ b )
Identification of parameters
Strategy
1
Learning phase : minimize with respect to ξ = (θ c , θ 1 b . . . , θ b 69 ) ∈ R 353 ξ ˆ = arg min
ξ J (ξ) = X
i=1...69
J i (θ c , θ b i ) + λkξk 2 ,
We obtain θ ˆ c ∈ R 8 which is used for the prediction
I
Optimization : interior points method (fminc, MATLAB)
I
Moderate computation time (computer with 20 processors Xeon E5-2660-v2)
2
Prediction/validation phase at T = 4 days : minimize with respect to θ i b ∈ R 5 θ ˆ i b = arg min
θ
bJ i T ( θ ˆ c , θ b ), i = 70 . . . 108
J i T ( θ ˆ c , θ b ) = X
k=2
t
k≤T
(t k i − t k−1 i )(D k i − d i k ( θ ˆ c , θ b )) 2
Results
Learning on batchs #1 to #69
Kinetic parameters θ ˆ c
Y X
10, 58 Y X
20, 12
µ 0 0, 29
µ max 1 3, 63 µ max 2 2, 67 K S
11, 02 K S
23, 45 K I 1, 44
0 200 400 600 800 1000 1200 1400
true BMP (NmL) 0
200 400 600 800 1000 1200
predicted BMP (NmL)
Model-3
Results
Learning, primary sludge
0 5 10 15
t (days) 0
100 200 300 400
flow (NmL/d)
Model-3 - batch n°5 (Manip 20130717, cell=8)
0 5 10 15
t (days) 0
500 1000
Vol (NmL))
Model-3 - batch n°5 (Manip 20130717, cell=8)
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°5 (Manip 20130717, cell=8)
S0 S1 S2
0 5 10 15
t (days) 0
5 10 15
g.COD/L
Model-3 - batch n°5 (Manip 20130717, cell=8) X1 X2
Results
Learning, nitrification sludge
0 5 10 15
t (days) 0
100 200 300 400
flow (NmL/d)
Model-3 - batch n°64 (Manip 20140116, cell=10)
0 5 10 15
t (days) 0
500 1000
Vol (NmL))
Model-3 - batch n°64 (Manip 20140116, cell=10)
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°64 (Manip 20140116, cell=10)
S0 S1 S2
0 5 10 15
t (days) 0
2 4 6 8
g.COD/L
Model-3 - batch n°64 (Manip 20140116, cell=10) X1 X2
Results
Prediction/validation on batchs #70 to #108 at T = 4 days
0 200 400 600 800 1000 1200 1400
true BMP (NmL) 0
200 400 600 800 1000 1200
predicted BMP (NmL)
Model-3
Results
Prediction, floated sludge
0 5 10 15
t (days) 0
100 200 300 400
flow (NmL/d)
Model-3 - batch n°92 (Manip 20140305, cell=11)
0 5 10 15
t (days) 0
500 1000
Vol (NmL))
Model-3 - batch n°92 (Manip 20140305, cell=11)
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°92 (Manip 20140305, cell=11)
S0 S1 S2
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°92 (Manip 20140305, cell=11) X1 X2
Results
Prediction, thickened sludge
0 5 10 15
t (days) 0
100 200 300 400
flow (NmL/d)
Model-3 - batch n°106 (Manip 20140407, cell=13)
0 5 10 15
t (days) 0
500 1000
Vol (NmL))
Model-3 - batch n°106 (Manip 20140407, cell=13)
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°106 (Manip 20140407, cell=13)
S0 S1 S2
0 5 10 15
t (days) 0
5 10
g.COD/L
Model-3 - batch n°106 (Manip 20140407, cell=13) X1 X2
Trends and conclusions
Results :
I
Well fitted kinetics in learning phase and good prediction of BMP at 4 days
I
Ratios of S
0, S
1, S
2seem to be interpretable
Planned improvements :
I
Theoretical study of identifiability of parameters in learning phase
I
DBO and VSS measurements should be taken into account
I
Coupling between triplicates has to be considered
I
Confidence intervals should be computed for the predicted BMP
I