Automatic history matching system and method for an oil reservoir based on transfer learning
The present invention relates to an automatic history matching system for an oil reservoir based on transfer learning, comprising a data reading module, a population reinitializing module, an optimization module, a simulated calculation module, a comparative judgment module and an output module, wherein the data reading module reads an optimized result of an existing oil reservoir, outputs the optimized result to the population reinitializing module, obtains an initial population of a new oil reservoir by calculation and outputs the initial population to the optimization module.
1 . An automatic history matching method for an oil reservoir based on transfer learning, adopting an automatic history matching system for an oil reservoir for improving the accuracy of reservoir forecasts in oil engineering, characterized in that the method comprises the steps of:
S1, reading observed data and an optimized result of an old model;
S2, performing population reinitializing processing on the optimized result of the old model to generate a reinitialized population; and
S3, performing evolutionary optimization calculation and simulated calculation by using the reinitialized population, predicting oil reservoir dynamically according to oil reservoir numerical simulator; characterized in that S1 comprises the specific steps of: assuming that the current moment is the moment (t+1), wherein the history matching model used by the oil reservoir at present is a t moment model, and as the oil reservoir at the moment (t+1) changes or the oil reservoir model is added with novel data, it is necessary to construct a history matching model for a current situation of the oil reservoir at the moment (t+1); reading the observed data of the oil reservoir at the moment (t+1) and providing preparation to judge whether the optimized result meets the requirement; and reading the optimized results of the history matching models at the moment (t−2), the moment (t−1) and the moment t, and constructing the initial population for the optimization process at the moment (t+1) by using its useful information for reference;
wherein the automatic history matching method is based on transfer learning, obtaining simulated oil production rates and liquid production rates for predicting future production status in oil engineering, for improving the accuracy of reservoir forecasts in oil engineering.
2 . The automatic history matching method for an oil reservoir based on transfer learning according to claim 1 , characterized in that S2 comprises the specific steps of: processing the optimized result of the history matching model of the oil reservoir at the moment t by using a reinitializing policy with directivity and random change based on transfer learning; extracting experience used as reference thereof; and constructing an initial population of a case at the moment (t+1).
3 . The automatic history matching method for an oil reservoir based on transfer learning according to claim 1 , characterized in that S3 comprises the specific steps of: setting an optimized target function that meets a specific requirement of the case at the moment (t+1), specific optimization indexes comprising, for example, oil production, water production and water content, and a formula (1) giving a loss function of the history matching model of a reservoir:
{ M =Σ( Q obs− Q cal)2} (1)
where M is an unmatched value, Qobs is true observed data of the oil reservoir, and Qcal is simulated calculation data by using the population obtained in the S2 as the initial population in the optimization process of the history matching model of the oil reservoir at the moment (t+1), using PSO or NSGA-III evolutionary optimization algorithm to optimize the history matching model of the oil reservoir at the moment (t+1) and adjusting the parameter thereof to obtain the optimized result at the moment (t+1); and
putting the optimized result into a reservoir numerical simulation for simulated calculation, wherein if an error between the observed data and the simulated data meets the requirement, an effective effect may be outputted, and on the contrary, iterative calculation is performed continuously.
4 . The automatic history matching method for an oil reservoir based on transfer learning according to claim 1 , characterized in that S2 specifically comprises: S21, controlled translation
calculating centroids of the optimum solution sets at three moments: C(t−2), C(t−1) and C(t) respectively according to optimum solution sets at the moment (t−2), the moment (t−1) and the movement t, making C(t)−C(t−1) to obtain a vector d(t), and making C(t−1)−C(t−2) to obtain a vector d(t−1); making C(t−1)+d(t+1) to obtain a vector b, defining an included angle between b and d(t) as cos−1cd, and calculating cd and further calculating a translation vector cdd(t), wherein all solutions in the optimum solution set at the moment t, i.e., all individuals and translation vectors in the optimized population at the moment t are added to reposition according to amplitude and direction of the translation vectors;
tracking movement of the optimum solution sets at the moment (t−2), the moment (t−1) and the moment t, thereby providing experience of being transferred to the optimum solution set that predicts a new problem at the moment (t+1); first, calculating movement of the optimized population individual at the moment t, wherein as shown in a formula (2), C(t) and C(t−1) are centroids of the population individual at the moment t and the moment (t−1);
dt=C ( t )− C ( t− 1) (2)
a predicted value C(t+1) is a centroid of the population individual in the optimums solution set at the moment (t+1), wherein a calculation method thereof is as shown in a formula (3),
C
^
(
t
+
1
)
=
C
(
t
)
+
c
d
d
(
t
)
c
d
=
max
{
min
{
1
,
d
(
t
)
.
d
(
t
-
1
)
d
(
t
)
.
d
(
t
-
1
)
}
,
0
}
(
3
)
where 0≤cd≤1 represents reliability that d(t) is taken as an evaluated value of d(t+1); as the two are usually different in both amplitude and direction, precise calculation may not be performed on the true d(t+1) in the stage; analyzing a predicted vale of d(t−1) to C(t) after moment change; moving C(t−1) directly with an accurate value close to C(t) according to direction and amplitude of d(t−1), wherein C(t−1) arrives a wrong point X; therefore, ∥b∥/∥d(t−1)∥ may calculate prediction reliability of d(t−1) to C(t), and the reliability is used to predict translation of C(t) along a direction of d(t) to obtain a distance of C(t+1);
S22, directional variation
performing random variation on all the population individuals processed in the S21 in a direction of the translation vector cdd(t), the change intensity thereof being amplitude of the translation vector multiplied with an independent random number vector N(1, σd2) sampled in normal distribution, wherein direction change is adopted as a specific form of random change to improve diversity of population obtained as the population translates along the direction of d(t), specifically as shown in a formula (4):
yj=xj *( t )+ cdd ( t )⊗ N (1,σ d 2), j= 1,2, . . . , Ns (4)
where yj is a population subject to controlled translation and directional variation, N(1, σd2) is an independent random number vector sample in normal distribution, a mean value thereof being 1, and a standard deviation being od; Ns is a size of population, and xj*(t) is the jth individual in the old optimum solution set at the moment t; under a default condition, the standard deviation is 1, and if yj is not in a search space, it is only needed to reposition yi to the nearest point in the search space; and in the repositioning process, whether each element of yj falls within the search space is judged, and if not, it is modified to an upper/lower bound value corresponding to the element; and
S23, random variation
applying completely random variable to all the population individuals processed in S22 to form an initial population of a new example (an example at the moment (t+1), the amplitude of random variation being determined by a mean value dpw of a distance between paired solutions at the moment (t−1) and the moment t,
wherein each population goes through uniform random change to generate white noise, intensity of which being determined by amplitude of random change; first, converting the optimum solution set POS(t−1) of the example at the moment (t−1) into C(t−1) and moving the same to be superposed with the centroid of C(t); and
then finding the nearest corresponding solution in POS(t−1) from a random solution in POS(t−1), matching two corresponding solutions and shielding the same to prevent from being selected by any other solution in POS(t) till all solutions in the two solution sets are paired, wherein an average distance between all the paired solutions is defined as dpw, which represents intensity of random variation;
σ r=crdpw
xj new ˜ N ( yj,σr 2, L,U ), j= 1,2, . . . , Ns (5)
where cr is a self defined value, xjnew is an initial population processed by the reinitializing policy with directivity and random change based on transfer learning at the moment (t+1), used for optimizing the history matching model of the oil reservoir at the moment (t+1), which is generated by using truncated normal distribution, wherein a center thereof is yj, a standard deviation is or, a lower bound is a lower bound (L) of the search space and an upper bound is an upper bound (U) of the search space.
5 . The automatic history matching method for an oil reservoir based on transfer learning according to claim 4 , characterized in that: the method for calculating the initial population of the oil reservoir model at the moment (t+1) comprises the specific steps of:
calculating a translation vector according to a formula (2);
calculating cd according to a formula (3);
calculating yj according to a formula (4), and modifying the same according to the method in the S22 if it is not in the search space; and
calculating a population reinitializing result according to a formula (5).