See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/371067791

Oil production forecast using Fuzzy Time Series and Artificial Neural Network
techniques

Conference Paper · January 2023

DOI: 10.3997/2214-4609.202310688

CITATION READS

1 22

2 authors:

Priyanshu Kumar Singh Nilanjan Pal

Indian Institute of Petroleum and Energy, Visakhapatnam Indian Institute of Petroleum and Energy
2 PUBLICATIONS 1 CITATION 58 PUBLICATIONS 2,966 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Priyanshu Kumar Singh on 28 December 2023.

The user has requested enhancement of the downloaded file.

Oil production forecast using Fuzzy Time Series and Artificial Neural Network
techniques
P.K. Singh1, N. Pal1

1
Indian Institute Of Petroleum & Energy

Summary
Accurate oil production forecasting is important for both companies and governments to make informed
decisions about the development and management of oil resources. A variety of methods and techniques
are available to predict future oil production, each method has its own advantages and disadvantages.
Oil production forecasting can be efficiently performed by Fuzzy Time Series (FTS) technique and by
Artificial Neural Network (ANN). Non-stationarity of data is a problem that is hard to deal with while
performing oil production forecasts. This paper proposes a new approach using FTS to predict future
oil production. This model can further be improved by incorporating this into hybrid systems. The
purpose of this study is to provide a brief detail about the methodology involved in the FTS and ANN
techniques to predict future oil production. Noise reduction and normalization are important aspects that
cannot be ignored when it comes to the improvement of the accuracy of the model. Different noise
reduction and data normalization technique are tested. Results obtained indicate that the FTS model
outperforms the ANN model.

84th EAGE Annual Conference & Exhibition

Oil production forecast using Fuzzy Time Series and Artificial Neural Network techniques

Introduction

Oil production forecast aid the oil companies, governments, and other stakeholders to determine the
most cost-effective and efficient ways to extract oil economically. There are various techniques and
methods for the oil production forecasting, such as: Decline Curve Analysis (DCA), Material Balance,
Statistical methods, ML models, Fuzzy Logic etc. The traditional Statistical Time Series Forecast (TSF)
methods include Auto Regressive Integrated Moving Average (ARIMA), Seasonal Auto Regressive
Integrated Moving Average (SARIMA) etc. In this paper the functionality of Simple-ANN and Fuzzy
Time Series (FTS) technique has been explored. The data set obtained for this study is from the Volve
Field well in the North Sea. Different forecasting methods have their own set of assumptions,
methodologies. For example, while performing DCA the analysis of past production data is done, and
accordingly using equation, that considers the decline rate and remain reserves, can be used for
forecasting the future productions. Another method, the Material Balance involves estimating the
amount of oil in place and the Recovery Factor (RF). This method can estimate the total amount of oil
that can be recovered from a given reservoir. Similarly, there is one more method called Volumetric
Analysis. This method considers the porosity and saturation of rock, instead of RF. Statistical method
like ARIMA, is a generalization of the simple exponential smoothing model and is used to forecast a
variety of time series data. SARIMA is an extended version of the ARIMA model. The only difference
between ARIMA and SARIMA is that SARIMA allows the modelling of seasonality in the data. When
using statistical method for forecasting for oil and gas production, long historical data is required, and
historical data is non-stationary kind of data. It is very important to consider a model that can handle
non-stationary data for the better prediction. ARIMA and SARIMA can handle most of the non-
stationary data, so it is widely used method for predictions. Alternatively, Neural Networks have shown
great potential in the time-series forecasting. Like the statistical method, historical data is required to
train the neural networks. FUZZY Logic, which is an unexplored area, has a great potential for time
series forecast and with a very less computational cost and less computational time. FTS is the other
aspect which is discussed in this paper.

Method and Theory:-

Field description – As mentioned earlier, the data set obtained for this study is from the Volve Field
well in the North Sea. Field description is provided in detail in (1). Briefly, the field started in
production in February 2008. Volve was operated by Equinor along with other partners ExxonMobil
E&P Norway and Bayerngas Norge. In September 2016 the field was decommissioned. In 2018, to
encourage study research and development, Equinor together with the other Volve Licence partners
decided to disclose all subsurface and operating data from the Volve Field. The Volve field data
available is of almost eight years right from beginning of production in 2008 till 2016.

Data Preparation- Quality and structure of data can have significant impact on the performance of the
model. So, it becomes important to consider data pre-processing an important step before feeding the
data into the model. Noisy data can negatively impact the performance of the model. Various noise
filtering technique were tested, such as – Rolling Mean, Exponential Weighted Moving Average
(EWM), Savitzky-Golay Filter (Savgol Filter). Graph obtained by applying Savgol filter is
represented in Figure 3. Plot in blue colour represent the training data and the plot in orange colour
represents the test data. After removing noise, the data was normalized using Min-Max Scalar for
training purpose. Figure 4 - 5 represents the normalized data on scale of 0 to 1.

Fuzzy-logic – Fuzzy logic was first proposed by Zadeh in 1965 (4). Classical set theory also known as
crisp set theory follows the rule that: either element is in OR out of the set, the membership of an
element is a boolean value, either 0 or 1. But according to Fuzzy set theory, a certain element may
belong and simultaneously do not belong to the same set at certain levels. Each element is given a
membership value in the interval [0,1] and the value represents the membership of the element towards
the set.

84th EAGE Annual Conference & Exhibition

 Figure 1 Original data Figure 2 Savgol Filter

Figure 3 EWM (Normalized) Figure 4 Savgol Filter(Normalized)

Methodology for FTS prediction:

Since noise reduction and normalization is performed beforehand, the rest of the steps in Fuzzy Time
series forecasting is as follows:

1) Fuzzification of the data- Partition scheme and the number of partitions is defined, after which
the elements of numerical time series are replaced by the fuzzy sets. Linguistic variable
(Universe of Discourse Partitioning) is generated for the fuzzy sets. GridPartition is used to
create partitions. Figure 8 represents the partition of linguistic variable. Equation (i) represents
the generated Linguistic variable Ã.

à = {A0, A1, A2, A3, A4…………………A34} (i)

2) Fuzzy rule generation: Based on the linguistic variable, the input data is fuzzified and then the
fuzzified data is used to train the model i.e., creation of fuzzy rules. The first few set of rules
generated after training the model is shown in Figure 5.

Figure 5 Fuzzy rules

3) Forecasting: Given the numerical value x(t), at time t, and the value x(t+1) is predicted. To
predict x(t+1), the input value must be fuzzified. For the fuzzified input find the compatible
rules from the fuzzy rules table obtained after the training process. The consequence of the rules
will be the fuzzy forecast for time (t+1). Now this fuzzy forecast is defuzzied to get the actual
predicted value. Center of mass method is applied for defuzzification. Figure 6 briefly represents
chronological order of the forecasting process.

Figure 6 Forecasting Process

84th EAGE Annual Conference & Exhibition

Methodology for ANN:
Method for this study is similar to method mentioned in (1), but different hyper-parameters and data
pre-processing technique is applied. The model performance is evaluated using the MSLE (Mean
Squared Logarithmic Error). It can be calculated using Equation (i).

(ii)

The model is trained on the 60% data. The training data is divided into set of batches. Window size is
the size of input to the ANN. MSLE was monitored at each training and testing stage to evaluate the
model performance. The trained model is then tested for the rest 40% data.

Results and Discussion

FTS - The experimental analysis for FTS was carried out using Weighted High Order FTS
(WHOFTS) as the base model within the TimeVariant metamodel. This metamodel at each batch size
input, retrains its base model with window size inputs. Important hyperparameters for FTS method
are: Number of Partitions, Lag (Order) and Membership Function (MF). After many trial and error
experiments, the optimal value of Number of Partitions and Order was found to be 35, 2 respectively.
Figure 8 represents the partitioning of the linguistic variable. Triangular membership function is
applied for this experimental study. In Figure 7, the slow decrease of ACF graph clearly indicates the
non-stationarity of the data used in this experiment. Compared to FTS other methods find it difficult
to deal with non-stationarity. It must be noted that FTS method, compared to other traditional
methods, require very less computational time and less computational cost. FTS methods are useful in
situations where traditional methods may not be appropriate, such as when data is noise, uncertain or
non-linear. FTS can easily handle non-stationary data. This field of interest is currently under research
and development phase. More powerful and robust methods can be established by combining FTS
methods with ARIMA, Recurrent Neural Network (RNN), Support Vector Regression (SVR), etc.

Figure 7 ACF Plot Figure 8 Partitions

ANN – The experimental analysis of Simple-ANN was carried using the same pre-processed data.
Gradient descent optimization method was used. Unlike FTS model learning process, gradient descent
algorithms are sensitive to the scale of input features, it becomes necessary to normalize the input data
to stabilize the learning process. The ANN model consists of 2 hidden layer, Rectified Linear Unit
(ReLU) activation function, Stochastic Gradient Descent (SGD) optimizer, and MSLE loss function.
With appropriate window size and batch size the model outputs satisfactory result. Figure 9 and
Figure 10 represents the results obtained from ANN model and FTS model respectively.

Figure 9 ANN prediction Figure 10 FTS prediction

84th EAGE Annual Conference & Exhibition

 After tuning the hyperparameters of each model, it was found that the MSLE was least for FTS model.
As depicted in Table 1, the MSLE value for FTS model was 1.095 whereas the MSLE value for ANN
model was 1.30. MAE value for FTS model was 73.14 whereas for ANN model was 102.56. Both MAE
and MSLE is minimum for FTS model.

Table 1 Showing computed of losses for the two models.

FTS model ANN model
MSLE 1.09454 1.30022
MAE 73.14777 102.56737

Conclusions

A new method of using FTS to forecast oil production in the volve field was proposed in this paper, and
thereafter the model is compared with ANN model. Oil production forecasting can be performed using
both the methods. The results indicate that the FTS method provides more accurate results than Simple-
ANN model. Study also conclude that FTS model can significantly reduce the computational cost and
time as well. Only drawback of FTS model is that it can only be used for short-term forecasts. To Fuzzy
systems, FTS-ARIMA, FTS with Genetic Algorithms (GA), FTS with Support Vector Machine (SVM),
FTS with Ensemble methods etc. can be studied and tested. Complex models like Long-short term
memory (RNN-LSTM) can be used for long-term forecast, but due to complexity they are not easy to
handle. Also, RNN-LSTM cannot handle all kinds of non-stationarity. No single method alone can
perform forecasting very accurately. So, it is a good idea to implement more than one method together
or use hybrid systems. Further research in the field of Fuzzy system should aim at developing models
like use of RNN with fuzzy Logic, RNN-LSTM with fuzzy Logic or Bi-directional RNNs with fuzzy
logic etc.

Acknowledgements

The authors acknowledge that the data set for this study was obtained from the Equinor’s website. File
were downloaded from Azure Cloud, using AzCopy software.

References

1) George, Alexander "Predicting Oil Production Flow Rate Using Artificial Neural Networks -
The Volve Field Case." Paper presented at the SPE Nigeria Annual International Conference
and Exhibition, Lagos, Nigeria, August 2021.
2) Chen Shyi-Ming. Forecasting enrollments based on fuzzy time series. Fuzzy sets and systems,
vol. 81, number 3, pp. 311–319, 1996.
3) Silva, P. C. L. et al. pyFTS: Fuzzy Time Series for Python, v4.0. 0
4) Zadeh, L. A. Fuzzy sets. Information and Control 8 (3) 338–353, 1965.
5) Serrano Ardila VM, Maciel JN, Ledesma JJG, Ando Junior OH. Fuzzy Time Series Methods
Applied to (In)Direct Short-Term Photovoltaic Power Forecasting. Energies. 2022; 15(3):845.
6) Silva, P.; Severiano, C.A.; Alves, M.A.; Silva, R.; Cohen, M.W.; Guimarães, F.G.
Forecasting in non-stationary environments with fuzzy time series. Appl. Soft
Comput. 2020, 97, 106825.
7) Petrônio Silva. A short tutorial on Fuzzy Time Series, https://towardsdatascience.com/a-short-
tutorial-on-fuzzy-time-series-dcc6d4eb1b15
8) Palomero, L.; García, V.; Sánchez, J.S. Fuzzy-Based Time Series Forecasting and Modelling:
A Bibliometric Analysis. Appl. Sci. 2022, 12, 6894
9) Rahmanifard, Hamid, Gates, Ian, and Abdolmohsen Shabib Asl. "Comparison of Machine
Learning and Statistical Predictive Models for Production Time Series Forecasting in Tight
Oil Reservoirs." Paper presented at the SPE/AAPG/SEG Unconventional Resources
Technology Conference, Houston, Texas, USA, June 2022.

84th EAGE Annual Conference & Exhibition

View publication stats