It is of great significance
to evaluate and predict coalbed methane
(CBM) production for the exploitation and exploration of CBM. The
flow characteristics of gas and water are very complicated and important
in the process of CBM exploitation. In recent years, machine learning
has been introduced to analyze CBM well production and its influence
based on the historical production data. However, there are some problems
with the determination of hyperparameters in machine learning algorithms.
Some previous random forests (RF) models of CBM production prediction
were suitable for individual CBM wells, but for different types of
CBM wells, a large amount of time is needed to adjust the hyperparameters.
Therefore, a genetic algorithm (GA) was applied to optimize RF, and
a hybrid GA–RF algorithm was presented to solve this problem,
which can automatically adjust two important hyperparameters,
n
tree
and
m
try
, and
adapt different types of CBM wells. Meanwhile, the Pearson method
and RF were carried out in this work to analyze the data of CBM well
production to avoid multicollinearity caused by the improper selection
of the model’s independent variables. The importance and correlation
analysis of drainage control parameters, including casing pressure
(
P
c
), bottom-hole pressure (
P
b
), stroke frequency (
f
s
),
liquid column depth (
D
L
), daily decline
of bottom-hole pressure (
P
bd
), and daily
decline of casing pressure (
P
cd
) were
obtained. It was found that the casing pressure, bottom-hole pressure,
and stroke frequency had more effects on the gas production of CBM
wells than other drainage control parameters. Furthermore, the correlation
and importance order of the influencing factors were:
P
c
>
P
b
>
f
s
>
P
bd
>
P
cd
>
D
L
and
P
c
>
P
b
>
f
s
>
D
L
>
P
bd
>
P
cd
, respectively.
A
CBM production model based on the GA–RF algorithm was constructed
to study and predict the gas production of CBM wells in Qinshui Basin,
China. Compared with the production model based on RF, this model
can automatically optimize its hyperparameters to adapt to different
types of CBM wells, and the mean-square-error of the GA–RF
algorithm can be reduced by 40–60% than that of RF. 93% of
the training errors were less than 5%, and 89% of the prediction errors
were less than 10%. The GA–RF model can spot promptly the ma...