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

net/publication/378550404

A Comparative Analysis of TimeGPT and Time-LLM in Predicting ESP

Maintenance Needs in the Oil and Gas Sector

Article in International Journal of Computer Applications · February 2024

DOI: 10.0.20.0/ijca2024923426

CITATION READS

1 1,991

2 authors:

Abhay Dutt Paroha Aakash Chotrani

Schlumberger Limited Oracle Corporation
11 PUBLICATIONS 10 CITATIONS 9 PUBLICATIONS 5 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Aakash Chotrani on 28 February 2024.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

A Comparative Analysis of TimeGPT and Time-LLM in

Predicting ESP Maintenance Needs in the Oil and Gas
Sector

Abhay Dutt Paroha Aakash Chotrani

Birla Institute of Technology And Science Oracle
Pilani (BITS Pilani), India USA
USA

ABSTRACT often reactive or based on predetermined schedules, have proven

inadequate in addressing the unpredictability and complexity of
This research evaluates the application of advanced artificial in- ESP failures [16]. Consequently, the industry has focused on
telligence models, TimeGPT and Time-LLM, for predictive main- predictive maintenance, which leverages data analysis and fore-
tenance of Electrical Submersible Pumps (ESPs) in the upstream casting to anticipate maintenance needs before failures occur [31].
oil and gas industry. The study meticulously analyzes the models’ This shift towards predictive maintenance promises to enhance
proficiency in forecasting maintenance needs, aiming to augment operational efficiency and reduce environmental impact by mini-
operational efficiency and reduce unplanned downtimes. Utilizing mizing unplanned outages and optimizing resource utilization [24].
a dataset rich in essential operational parameters, the comparative
analysis reveals TimeGPT’s marginally superior performance, with In predictive maintenance, the advent of Artificial Intelligence (AI)
an accuracy of 95.2%, precision of 92.8%, recall of 94.1%, and an and machine learning (ML) has introduced transformative potential
AUC-ROC of 0.971. In contrast, Time-LLM achieves an accuracy [8], particularly through advanced time series forecasting models.
of 93.6%, precision of 90.5%, recall of 91.2%, and an AUC-ROC Two such models at the forefront of this technological wave are
of 0.957. Both models effectively identify critical indicators of ESP TimeGPT [12] and Time-LLM [17], epitomizing deep learning
health, aligning with established industry knowledge. The integra- techniques’ integration in predictive analytics. These models are
tion challenges of these AI models into existing industrial setups engineered to analyze and interpret complex, time-dependent
are discussed, underscoring the necessity for high-quality data and data [12, 17], a characteristic intrinsic to ESP operations [4]. The
system compatibility. The study suggests future research directions, predictive capabilities of TimeGPT and Time-LLM are due to
emphasizing model refinement, economic impact assessment, and the ability to detect patterns and anomalies in extensive datasets,
AI technology’s ethical and environmental considerations. This re- encompassing variables such as operational parameters, historical
search provides significant insights into the use of AI in industrial performance data, and environmental factors. This analysis is
maintenance, marking a stride toward more proactive and data- critical in predicting ESP maintenance needs, aiming to preempt
driven operational strategies in the oil and gas sector. equipment failures and optimize maintenance schedules [3].
However, the comparative efficacy of these models in the specific
Keywords context of ESP maintenance remains an area ripe for exploration,
Predictive Maintenance, Artificial Intelligence (AI), TimeGPT with significant implications for operational efficiency, cost
Model, Time-LLM Model, Electrical Submersible Pumps (ESPs), reduction, and strategic planning in the oil and gas sector.
Oil and Gas Industry, Operational Efficiency, Machine Learning
Algorithms, Data Analytics in Energy Sector, Maintenance Strat- This research aims to revolutionize maintenance practices within
egy Optimization the upstream oil and gas industry, where operational efficiency and
equipment reliability are paramount [5]. The current field, domi-
1. INTRODUCTION nated by conventional maintenance strategies, often leads to costly
and unforeseen equipment downtime, particularly in the case of
The upstream oil and gas industry, characterized by its intensive ESPs [30]. These challenges underscore the necessity for a more
resource utilization and complex operational dynamics, has proactive and data-driven approach to maintenance. Prior works
consistently sought innovative methods to enhance efficiency and have increasingly centered on utilizing the capabilities of machine
reliability, particularly in equipment maintenance [9]. Within this learning and deep learning to predict maintenance requirements.
context, Electrical Submersible Pumps (ESPs) emerge as critical Key studies have particularly emphasized the role of AI in moni-
components, pivotal to the extraction process, yet susceptible to toring and predicting the needs of ESPs, vital components in oil ex-
operational wear and failures [1] that can lead to significant down- traction operations. However, a comprehensive comparative analy-
time and financial losses [11]. Traditional maintenance strategies,

1
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

sis of these models, specifically tailored to the demands and intri- volving routine inspections and scheduled servicing based on es-
cacies of the oil and gas sector, remains conspicuously absent from timated equipment lifespans and operational demands [30]. Pre-
the literature. This gap underscores the need for an in-depth evalu- ventive maintenance represented a more systematic approach to
ation of these models’ effectiveness, a task this research endeavors equipment care, aiming to avoid failures before they happened [2].
to fulfill. The paper aims to address the following Research Objec- However, despite its proactive nature, this strategy was not with-
tives: out drawbacks. It often led to unnecessary maintenance activities,
operational inefficiencies, and increased costs [2]. Over time, the
(1) To thoroughly evaluate the accuracy and reliability of limitations of preventive maintenance became clear, particularly its
TimeGPT and Time-LLM in predicting ESP maintenance lack of flexibility and inability to adapt to the specific conditions
needs. and usage patterns of equipment [18].
(2) To assess both models’ computational efficiency and scalabil- This need for a more dynamic and responsive maintenance ap-
ity in handling datasets specific to ESP operations. proach led to the emergence and gradual adoption of predictive
(3) To explore the integration feasibility of TimeGPT and Time- maintenance in the oil and gas industry [33]. Predictive mainte-
LLM within the existing technological framework of the oil nance, a strategy that employs real-time data analysis and condi-
and gas industry. tion monitoring tools to predict when maintenance should be per-
formed, significantly improved the optimization of maintenance
The exploration and comparative analysis of advanced AI models schedules [6]. Predictive maintenance began transforming the in-
like TimeGPT and Time-LLM for predictive maintenance becomes dustry’s equipment maintenance approach by the early 21st cen-
not just relevant but essential. These models promise to transform tury with advancements in sensor technology and data analyt-
maintenance strategies from reactive to predictive, enhancing op- ics capabilities [21]. Unlike preventive maintenance, which relies
erational efficiency, minimizing downtime, and reducing costs. By on generalized schedules, predictive maintenance utilizes actual
accurately forecasting ESP maintenance needs, these AI-driven ap- equipment performance data and operational conditions to forecast
proaches can lead to more informed decision-making, better re- maintenance needs [26]. This shift improved operational efficiency
source allocation, and a more sustainable and profitable operation and significantly reduced unplanned downtimes and maintenance-
in the oil and gas sector. This study, therefore, seeks to bridge the related costs. Predictive maintenance’s ability to identify potential
gap between advanced AI technology and practical, real-world ap- equipment failures before they occur was particularly beneficial in
plication in a critical industry, providing valuable insights and guid- managing critical equipment such as ESPs [26].
ing future advancements in predictive maintenance practices. The ESPs, essential in the extraction process in the oil and gas in-
paper adds the following contributions in the current literature: dustry, particularly in wells with low natural pressure, present
(1) This analysis contributes to the literature by providing empir- unique maintenance challenges [20]. Operating in harsh environ-
ical evidence on the predictive capabilities of these models, ments characterized by high pressures and temperatures, ESPs are
aiding in the refinement of maintenance strategies within the susceptible to frequent wear and tear, making their maintenance a
oil and gas sector. critical aspect of upstream operations [28]. The failure of an ESP
can lead to substantial production losses, safety risks, and envi-
(2) This objective contributes insights into the practicality of de-
ronmental concerns. Therefore, applying predictive maintenance to
ploying these models in real-world industry scenarios, influ-
ESPs became a focal point in the industry’s efforts to enhance oper-
encing decisions on resource allocation and infrastructural de-
ational reliability and efficiency [29]. Deploying sensors and mon-
velopment.
itoring systems capable of withstanding the ESPs’ operational en-
(3) This investigation enriches understanding of the challenges vironment provided the necessary data for predictive analysis [15].
and prerequisites for implementing advanced AI models in ex- This data, which includes metrics such as temperature, vibration,
isting workflows, thereby guiding future innovations in AI in- and pressure, is crucial in identifying signs of potential wear or fail-
tegration for predictive maintenance. ure, enabling timely maintenance interventions before catastrophic
The rest of the paper is organized in the following manner. Section failures occur.
2 provides a detailed literature review of this field, from the his- AI and ML have revolutionized the field of predictive maintenance,
torical overview to the progress. Section 3 provides the proposed bringing about a paradigm shift in how data is processed and an-
methodology, materials, and methods of TimeGPT and Time-LLM. alyzed [27]. AI and ML algorithms can handle large volumes of
section 4 provides the experiment results, whereas section 5 dis- complex data at speeds and accuracy levels beyond human capabil-
cusses those results and provides a comparative model overview. ities. In ESP maintenance, these technologies analyze operational
Finally, section 6 concludes this paper with future recommenda- data, identifying patterns and anomalies that might indicate im-
tions. pending failures [13]. This unique and sophisticated analysis level
was previously unattainable with traditional data analytics meth-
ods. AI and ML have thus enabled a more precise and predictive ap-
2. LITERATURE REVIEW
proach to ESP maintenance, minimizing downtime, reducing main-
Maintenance practices in the upstream oil and gas industry, charac- tenance costs, and improving overall operational efficiency [2].
terized by intensive resource utilization and complex operational Recent developments in AI, notably advanced time series forecast-
dynamics, have been marked by significant evolutions [19]. Ini- ing models such as TimeGPT and Time-LLM, have further pushed
tially dominated by reactive maintenance strategies, the industry’s the boundaries of predictive maintenance [12, 17]. These models
approach addressed equipment failures as they occurred [32]. This use deep learning techniques to analyze and interpret complex,
method, while straightforward, often resulted in substantial op- time-dependent data, a characteristic intrinsic to ESP operations.
erational downtimes and financial burdens [14]. As the industry Their ability to detect subtle patterns and anomalies in extensive
evolved, so did the understanding of the inefficiencies and limi- datasets, encompassing variables such as operational parameters,
tations inherent in a purely reactive approach [25]. This realiza- historical performance data, and environmental factors, has proven
tion prompted a shift towards preventive maintenance strategies in- crucial in predicting ESP maintenance needs. However, the practi-

2
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

cal application of these advanced models in the oil and gas industry 3.1 Data Collection and Preparation
is not without challenges [10]. Integrating these models into exist-
ing technological frameworks requires careful consideration of var- The dataset utilized in this study is a comprehensive collec-
ious factors, including the need for skilled personnel, infrastructure tion of operational data from ESP in the oil and gas sector,
upgrades, and compatibility with existing data systems. spanning 58,978 records. It encompasses various variables cru-
The evolution of maintenance practices in the upstream oil and gas cial for ESP maintenance and performance monitoring. Parame-
industry, particularly ESP maintenance, has been marked by a con- ters such as ’CURRENT’, ’PRESS DESC’, ’FREQUENCY’, and
tinuous search for greater efficiency and reliability [22]. The shift ’TEMP INT’ provide essential electrical and pressure data, while
from reactive to preventive, and eventually to predictive mainte- ’VIBRATION’ metrics offer insights into potential mechanical in-
nance, reflects the industry’s response to its operational challenges tegrity issues. Production metrics, including ’BFPD’, ’BOPD’, and
and technological advancements. The integration of AI and ML in ’BWPD’, reflect the output performance and, in conjunction with
predictive maintenance, exemplified by models like TimeGPT and ’AMPERAGE’ and ’WHP(PSI)’, serve as indicators of the ESPs’
Time-LLM, represents this evolution’s current frontier, offering un- operational conditions. The dataset is enriched with a temporal di-
precedented equipment maintenance and management capabilities. mension, furnishing a framework for time-series analyses vital for
As the industry continues to navigate its complex operational en- predictive maintenance modeling. Including a ’FAILURE’ label,
vironment, the role of advanced predictive maintenance strategies denoting the occurrence of a pump failure, establishes a founda-
powered by AI and ML is likely to become increasingly central, tion for supervised machine learning, enabling training models like
driving improvements in operational efficiency, cost management, TimeGPT and Time-LLM to forecast maintenance needs equip-
and equipment reliability. ment failures, thereby enhancing operational efficiency and reduc-
ing downtime.
3. MATERIALS AND METHODS 3.2 Model Development and Training
The methodology of this study is designed to evaluate the effec- The methodology’s core involves developing and training the
tiveness of TimeGPT and Time-LLM models in predicting mainte- TimeGPT and Time-LLM models. Each model is configured and
nance needs for ESPs in the upstream oil and gas industry. The trained separately using the collected ESP operational data.
research approach integrates data collection, model training and
testing, and comparative analysis to assess these models’ accuracy, 3.2.1 TimeGPT. TimeGPT employs a transformative architec-
computational efficiency, and practical applicability. The overall ar- ture derivative of the Generative Pretrained Transformer (GPT) tai-
chitecture of the methodology is provided in the figure 1. lored for temporal data analysis. Its structure is composed of mul-
tiple layers of self-attention mechanisms, each layer consisting of
two sub-layers: the multi-head self-attention mechanism and the
position-wise feed-forward network. The architecture diagram of
TimeGPT is provided in the figure 3. The architecture is predicated
on the following key equations:- Self-attention can be defined as:
 ′ ′T 
QK
SA(Q′ , K ′ , V ′ ) = softmax √ V′ (1)
dk ′
Multi-head attention:
M HA(Q′ , K ′ , V ′ ) = Concat(head′1 , ..., head′h )W ′O (2)

where head′i = SA(Q′ Wi′Q , K ′ Wi′K , V ′ Wi′V ) (3)

Position-wise Feed-Forward Networks:
F F N (x′ ) = max(0, x′ W ′1 + b′1 )W ′2 + b′2 (4)
Layer Normalization and Residual Connection:
x′′ = LayerNorm(x′ + Sublayer(x′ )) (5)

where Sublayer(x′ ) is an operation applied to x′ (6)

Each attention head captures different aspects of the data, allowing
the model to focus on different positions within the input sequence.
Normalization and residual connections are included at each sub-
layer, following the equations:
LayerN orm(x + Sublayer(x)) (7)
where Sublayer(x) is the function implemented by the sub-layer
itself. TimeGPT’s training employs a robust optimization strategy
using the Adam optimizer, with a custom learning rate scheduler
that increases the learning rate linearly for the first warm-up steps
and then decreases it proportionally to the inverse square root of
Fig. 1. The architecture diagram of overall Methodology the step number.

3
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Fig. 2. The architecture diagram of TimeGPT, image source [23]

This experiment utilized a TimeGPT model with 12 layers, 768 multi-head self-attention mechanism, reflecting the intricacies of
hidden units, 12 heads, and a feed-forward filter size 3072. The the input data. The model underwent training on the ESP dataset
model was trained on the ESP dataset with a batch size of 64, a over 50 epochs with a batch size 32. A learning rate of 1e-4 was se-
learning rate of 2.5e-4, and a warm-up period of 10,000 steps to lected, with a scheduler reducing this rate by half every 10 epochs
ensure gradual and stable learning. The model’s predictions were once a plateau in validation loss was detected. To counteract over-
then evaluated against a hold-out validation set, assessing the ac- fitting, a dropout rate of 0.1 was employed during training. The
curacy and mean squared error to quantify the model’s predictive AdamW optimizer with weight decay was used to refine the model
performance. This experimental setup aims to reflect real-world op- parameters for optimization.
erational conditions and provide insights into the feasibility of de- Validation was performed on a separate subset of the data, where
ploying TimeGPT for predictive maintenance within the oil and gas the model’s predictions were measured for accuracy, precision, re-
sector. call, and F1-score to evaluate its performance comprehensively.
The test set, drawn from a distinct time, ensured that the model’s
3.2.2 Time-LLM. Time-LLM, while conceptualized within the predictive power generalized well to unseen data. This rigorous ex-
same realm of predictive analytics as TimeGPT, distinguishes it- perimental setup was designed to closely simulate the operational
self through its architecture and operational methodology tailored conditions of the oil and gas industry’s ESP systems.
to time-series language modeling. Time-LLM integrates long short-
term memory (LSTM) networks with transformer models, capital- 3.3 Model Testing and Validation
izing on LSTM’s ability to retain information over extended periods
and the transformer’s efficient handling of dependencies. After training, both models will undergo a rigorous testing and vali-
The architecture combines the LSTM’s gated mechanism with dation process. A separate set of data, not used in the training phase,
the transformer’s multi-head self-attention and position-wise feed- will be utilized to evaluate the models’ predictive capabilities. This
forward networks. The following equations govern the LSTM com- phase assesses how accurately each model can predict ESP mainte-
ponent: Forget Gate: nance needs and potential failures. The performance of the models
will be measured using metrics such as accuracy, precision, recall,
ft′ = σ(Wf ′ · [h′t−1 , x′t ] + bf ′ ) (8) and F1-score.
Input Gate:
3.4 Comparative Analysis
i′t = σ(Wi′ · [h′t−1 , x′t ] + bi′ ) (9)
A comparative analysis of the TimeGPT and Time-LLM models
Cell Update: is conducted to determine their relative effectiveness. This analysis
will focus on predictive accuracy and evaluate each model’s com-
C̃t = ft′ ∗ C̃t−1 + i′ t ∗ tanh(W C ′ · [h′t−1 , x′ t] + bC ′ ) (10)
putational efficiency and scalability. Factors like processing time,
Output Gate: resource utilization, and ease of integration into existing mainte-
nance workflows will be considered.
o′t = σ(Wo′ · [h′t−1 , x′t ] + bo′ ) (11)
Hidden State Update: 3.5 Feasibility Assessment for Industry Integration

h′t = o′t ∗ tanh(C̃t ) (12) An essential aspect of the methodology is assessing the feasibility
of integrating TimeGPT and Time-LLM models into the existing
The architecture diagram of Time-LLM is provided in figure 3. For technological framework of the oil and gas industry. This assess-
Time-LLM, This experimental configuration was established with ment will involve consultations with industry experts, field engi-
an architecture comprising 10 LSTM layers, each with 256 hidden neers, and IT specialists to understand the practical challenges and
units, to capture the long-term dependencies characteristic of ESP prerequisites for implementing these AI models in real-world sce-
time-series data. The transformer section included 8 heads in the narios.

4
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Fig. 3. The architecture diagram of Time-LLM, image source [7]

4. RESULTS AND ANALYSIS sion during the training of neural networks. The initial loss values
for TimeGPT start at 4.78 for training and 9.43 for validation, sig-
4.1 Exploratory Data Analysis nifying the model’s initial unfamiliarity with the patterns within
The Correlation heatmap in figure 4 visually represents the cor- the data. As training progresses, both curves exhibit a smooth,
relation coefficients between different variables in the dataset. It quadratic decline towards their respective minimum values, 0.0032
helps identify potential relationships or dependencies among vari- for training and 0.0031 for validation, reflecting the model’s in-
ous operational parameters, such as how pump pressure, tempera- creasing proficiency in predicting ESP failures as it learns from the
ture, and vibration are interrelated. These strong correlations guide training data. For the Time-LLM model, the initial training loss
the focus on specific variables for further analysis and model fea- begins at a higher value of 8.671, suggesting a steeper learning
ture selection. The figure 5 shows the Boxplot of Pressure Descent curve when compared to TimeGPT. The validation loss for Time-
(PRESS DESC), which visualizes the distribution and variability LLM starts at 7.331, notably lower than its training counterpart,
of the Pressure Descent values. The presence of outliers can be potentially indicating a model that is initially more generalizable.
readily identified, which is essential for understanding the range However, as epochs advance, the training loss surpasses the vali-
and anomalies in pressure conditions. The central box represents dation loss, suggesting overfitting may occur. Ultimately, the train-
the interquartile range (IQR), with the line inside denoting the me- ing loss converges to a lower minimum value of 0.0022, compared
dian. Whiskers extend to show the range of the data, excluding out- to a slightly higher validation loss of 0.0048, hinting at the model
liers. The histogram for the ’CURRENT’ variable shows the dis- capturing the underlying trends in the data and retaining some sus-
tribution of electrical current values. The shape of the distribution, ceptibility to fluctuations in unseen data. These loss curves provide
whether normal or skewed, can provide insights into the typical valuable insights into the learning dynamics of the models. The
operating conditions of the pumps. Multiple peaks (if any) can in- presence of fluctuations and the convergence pattern indicate the
dicate different operational states or regimes. This scatter plot of models’ capabilities to assimilate complex temporal relationships
Amperage (PRESS INT vs. AMPERAGE) examines the relation- within the dataset, which is essential for the predictive maintenance
ship between Pressure Intake and Amperage. Any visible pattern of ESPs. The curves’ crossing points, where the validation loss dips
or trend can indicate how pressure changes affect the pumps’ elec- below the training loss, reflect moments where the model’s predic-
trical demand. Such a relationship is crucial in predictive mainte- tions align more closely with unseen data, a desirable attribute in
nance, as deviations from the normal pattern can signal potential a predictive model. These patterns underscore the iterative refine-
issues. ment of the models’ weights and parameters, guiding them towards
a state of optimal predictive performance.
4.2 Training Progress
The loss curves for both TimeGPT and Time-LLM are provided in
figure 8 and 9 to accurately represent the typical learning progres-

5
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Fig. 4. The Correlation Diagram of Features produced by applying random forest feature importance

Fig. 5. The Boxplot of Pressure Descent feature

Fig. 6. The histogram for the ’CURRENT’ variable

4.3 Results of Experiments
The analysis conducted in this study aimed to evaluate the efficacy binary outcome. The results presented here offer insights into the
of the TimeGPT and Time-LLM models in predicting maintenance predictive capabilities of these models.
requirements for ESPs in the oil and gas industry. The models were The TimeGPT model demonstrated notable proficiency in forecast-
trained and validated on a comprehensive dataset comprising vari- ing ESP failures. Upon evaluation, the model achieved an accu-
ous operational parameters, focusing on predicting the ’FAILURE’ racy of 95.2%, with a precision of 92.8% and a recall of 94.1%.

6
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Fig. 7. This scatter plot of Amperage (PRESS INT vs. AMPERAGE)

Fig. 9. The loss metric of Time-LLM

Fig. 8. The loss plot of TimeGPT

Fig. 10. The confusion metric of TimeGPT
The F1-score, which balances the precision and recall, was calcu-
lated at 93.4%. These metrics underscore the model’s ability to ac- than the TimeGPT model. The figure 13 shows the ROC curve of
curately identify potential failure events, while maintaining a low Time-LLM.
rate of false positives. The figure 10 shows the confusion metric
of TimeGPT. The area under the Receiver Operating Characteris- 4.4 Comparative Analysis
tic (ROC) curve (AUC-ROC), a measure of the model’s ability to
distinguish between the failure and non-failure classes, was 0.971. The TimeGPT model, with its intricate architecture inspired by the
This high AUC-ROC value indicates a strong discriminatory power Generative Pretrained Transformer, demonstrated marginally supe-
of the TimeGPT model in classifying the ESP conditions. The fig- rior performance over the Time-LLM model in key predictive met-
ure 11 shows the ROC curve of TimeGPT. rics such as accuracy, precision, recall, and the Area Under the Re-
The Time-LLM model also exhibited commendable performance, ceiver Operating Characteristic curve (AUC-ROC). These metrics
albeit slightly lower than TimeGPT. It achieved an accuracy of are crucial in predictive maintenance, where the cost of false pre-
93.6%, a precision of 90.5%, and a recall of 91.2%. The F1-score dictions – both false positives and negatives – can be substantial.
for Time-LLM was recorded at 90.8%. These results reflect the TimeGPT’s slightly higher scores suggest a more refined capabil-
model’s effectiveness in identifying failure events, albeit with a ity in correctly identifying potential ESP failures and, crucially, in
marginally higher rate of false negatives than TimeGPT. The fig- avoiding false alarms, which can lead to unnecessary maintenance
ure 12 shows the confusion metric of Time-LLM. actions and associated costs.
The AUC-ROC for the Time-LLM model was 0.957, indicating ro- Time-LLM, incorporating elements of both LSTM (Long Short-
bust classification capabilities, although marginally less effective Term Memory) and transformer models while slightly trailing be-

7
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Fig. 11. Roc Curve of TimeGPT Fig. 13. The confusion metric of Time-LLM

schedules, and enhanced overall operational efficiency. However,

this integration is not without challenges.
One of the key limitations in applying these models is the need for
a robust and continuous stream of high-quality data. These mod-
els’ performance heavily depends on the quantity and quality of
the data fed into them. Inconsistent or poor-quality data can lead
to inaccurate predictions, potentially undermining the effectiveness
of the maintenance strategy. Moreover, implementing these models
in real-world scenarios requires careful consideration of the exist-
ing technological infrastructure within the oil and gas companies.
Integration challenges may arise due to compatibility issues with
current data management systems, necessitating potentially costly
upgrades or modifications.
Another critical aspect is the ongoing maintenance and updating
of these models. As operational conditions and equipment char-
acteristics evolve, the models must be regularly updated and re-
trained to maintain their accuracy and relevance. This requirement
for continuous oversight and resource investment might be a con-
sideration for companies adopting these technologies. While the
TimeGPT model shows a slight edge in predictive performance,
both TimeGPT and Time-LLM present valuable tools in the arse-
nal of predictive maintenance strategies for the oil and gas industry.
Fig. 12. The confusion metric of Time-LLM
Their efficacy in identifying key failure indicators aligns well with
industry knowledge, enhancing their applicability in real-world set-
hind TimeGPT in these performance metrics, still exhibited com- tings. However, successful implementation would require address-
mendable predictive capabilities. This performance underscores ing challenges related to data quality, technological integration, and
the model’s potential utility in predictive maintenance, especially ongoing model management. Despite these challenges, the poten-
considering its different architectural approaches, which may of- tial benefits of improved operational efficiency and reduced main-
fer unique advantages in certain operational contexts. Both mod- tenance costs make a compelling case for adopting these advanced
els identified vibration, pressure descent, and temperature intake AI models.
as significant predictors of ESP failure. This alignment with ex-
isting industry knowledge about critical indicators of pump health 5. CONCLUSION
validates the models’ predictive relevance and enhances their cred-
ibility and potential for adoption in the field. The emphasis on The research on the effectiveness of TimeGPT and Time-LLM
these parameters underscores the importance of real-time monitor- models in predictive maintenance for ESPs in the oil and gas indus-
ing and data collection in the operational environment for effective try marks a significant stride in applying advanced artificial intel-
predictive maintenance. Both models’ high accuracy and reliabil- ligence techniques to real-world industrial challenges. This study’s
ity in predicting ESP failures indicate their significant potential to findings underscore the robust capabilities of these models in accu-
transform maintenance strategies within the oil and gas sector. In- rately predicting maintenance needs, thereby heralding a new era
tegrating AI-driven predictive models into operational workflows in the operational strategies of the oil and gas sector. The compar-
promises reduced unplanned downtimes, optimized maintenance ative analysis revealed a nuanced distinction in the performance of

8
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

the TimeGPT and Time-LLM models. While TimeGPT exhibited a [5] Sahab Ali. The concept of energy efficiency technologies in
marginal edge in accuracy, precision, recall, and AUC-ROC, Time- the upstream petroleum industry: A literature review. Master’s
LLM too, demonstrated commendable performance levels. Impor- thesis, Høgskolen i Molde-Vitenskapelig høgskole i logistikk,
tantly, both models identified critical operational parameters – vi- 2020.
bration, pressure descent, and temperature intake – as key indica- [6] Serkan Ayvaz and Koray Alpay. Predictive maintenance sys-
tors of ESP health, aligning with established industry knowledge. tem for production lines in manufacturing: A machine learn-
This alignment corroborates the models’ reliability and augments ing approach using iot data in real-time. Expert Systems with
their potential for practical application, offering a pathway to en- Applications, 173:114598, 2021.
hanced efficiency and reduced operational costs through proactive
[7] conceal. Time-llm: Time series forecasting by reprogram-
maintenance strategies.
ming larage language models. https://blog.csdn.net/
The implications of integrating such AI-driven predictive models
weixin_44965236/article/details/133739374, 2023.
into the oil and gas industry are profound. They signify a shift from
Accessed: 19-02-2024.
traditional reactive maintenance to a more data-informed, proactive
approach. However, this transition is not devoid of challenges. The [8] Maria Drakaki, Yannis L Karnavas, Ioannis A Tziafettas,
primary obstacle lies in the necessity for high-quality, consistent Vasilis Linardos, and Panagiotis Tzionas. Machine learning
operational data, which forms the backbone of any AI model’s ef- and deep learning based methods toward industry 4.0 predic-
fectiveness. Inconsistent or poor-quality data can significantly di- tive maintenance in induction motors: State of the art survey.
minish the predictive accuracy, potentially leading to misguided Journal of Industrial Engineering and Management (JIEM),
maintenance actions. Thus, there is an imperative for companies 15(1):31–57, 2022.
to bolster their data infrastructure, ensuring the continuous flow of [9] Olakunle Elijah, Pang Ai Ling, Sharul Kamal Abdul Rahim,
accurate and comprehensive data. Additionally, integrating these Tan Kim Geok, Agus Arsad, Evizal Abdul Kadir, Muslim
advanced models into existing technological frameworks poses its Abdurrahman, Radzuan Junin, Augustine Agi, and Moham-
own set of challenges. Issues of compatibility with current systems, mad Yasin Abdulfatah. A survey on industry 4.0 for the oil
scalability, and the cost of implementation must be judiciously ad- and gas industry: Upstream sector. IEEE Access, 9:144438–
dressed. To navigate these complexities, it is recommended that 144468, 2021.
companies engage in detailed feasibility studies and seek collab- [10] Emmanuel I Epelle and Dimitrios I Gerogiorgis. A review of
oration with experts in AI and data science. technological advances and open challenges for oil and gas
Looking ahead, the study opens numerous avenues for further re- drilling systems engineering. AIChE Journal, 66(4):e16842,
search. There is ample scope for refining these models to accom- 2020.
modate a wider spectrum of operational conditions and equipment [11] Sherif Fakher, Abdelaziz Khlaifat, M Enamul Hossain, and
types. Exploring incorporating additional data types, such as acous- Hashim Nameer. Rigorous review of electrical submersible
tic or seismic data, could further enhance the models’ predictive pump failure mechanisms and their mitigation measures.
capabilities. Beyond technical enhancements, an analysis of the Journal of Petroleum Exploration and Production Technol-
economic impact of implementing these AI models is vital. Un- ogy, 11:3799–3814, 2021.
derstanding the cost-benefit dynamics and the potential return on
investment would provide valuable insights for stakeholders in the [12] Azul Garza and Max Mergenthaler-Canseco. Timegpt-1.
oil and gas industry. Moreover, as research advances in this field, arXiv preprint arXiv:2310.03589, 2023.
it is imperative to consider the ethical and environmental implica- [13] Supriya Gupta, Luigi Saputelli, and Michael Nikolaou. Big
tions of deploying such technologies. Issues surrounding data pri- data analytics workflow to safeguard esp operations in real-
vacy, security, and the responsible use of AI need to be at the fore- time. In SPE Artificial Lift Conference and Exhibition-
front of this technological evolution. Furthermore, the environmen- Americas?, page D021S004R003. SPE, 2016.
tal impact, particularly regarding energy consumption and carbon [14] Dennis Harris, Mark Banman, and David Malone. Design and
footprint associated with these technologies, warrants careful con- qualification testing of esp cable to improve esp system run
sideration. life. In SPE Gulf Coast Section Electric Submersible Pumps
Symposium?, page D021S003R003. SPE, 2019.
6. REFERENCES [15] Nico Jansen Van Rensburg, Lisa Kamin, and Skip Davis.
Using machine learning-based predictive models to enable
[1] Ramez Abdalla, Denis Nikolaev, David Gönzi, Roman Man- preventative maintenance and prevent esp downtime. In Abu
asipov, Andreas Schweiger, and Michael Stundner. Deep in- Dhabi International Petroleum Exhibition and Conference,
sight into electrical submersible pump maintenance: A pre- page D021S043R004. SPE, 2019.
dictive approach with deep learning. In SPE Offshore Europe
[16] Veronica Jaramillo Jimenez, Noureddine Bouhmala, and
Conference and Exhibition, page D021S006R004. SPE, 2023.
Anne Haugen Gausdal. Developing a predictive maintenance
[2] Ramez Abdalla, Hanin Samara, Nelson Perozo, Carlos Paz model for vessel machinery. Journal of Ocean Engineering
Carvajal, and Philip Jaeger. Machine learning approach for and Science, 5(4):358–386, 2020.
predictive maintenance of the electrical submersible pumps [17] Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y
(esps). ACS omega, 7(21):17641–17651, 2022. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-
[3] Mohannad Abdelaziz, Rafael Lastra, and JJ Xiao. Esp data Fang Li, Shirui Pan, et al. Time-llm: Time series forecast-
analytics: Predicting failures for improved production perfor- ing by reprogramming large language models. arXiv preprint
mance. In Abu Dhabi international petroleum exhibition & arXiv:2310.01728, 2023.
conference. OnePetro, 2017. [18] Anne Kwong, Junaid Hussain Muzamal, and Zohaib Khan.
[4] Shaikha AlBallam. Applying machine learning models to di- Privacy pro: Spam calls detection using voice signature anal-
agnose failures in electrical submersible pumps. 2022. ysis and behavior-based filtering. In 2022 17th International

9
 International Journal of Computer Applications (0975 - 8887)
Volume 186 - No.8, February 2024

Conference on Emerging Technologies (ICET), pages 184– [32] Rama S Velmurugan and Tarun Dhingra. Maintenance strat-
189. IEEE, 2022. egy selection and its impact in maintenance function: A con-
[19] Paola Lara, Mario Sánchez, and Jorge Villalobos. Enterprise ceptual framework. International Journal of Operations &
modeling and operational technologies (ot) application in the Production Management, 35(12):1622–1661, 2015.
oil and gas industry. Journal of Industrial Information Inte- [33] Tiago Zonta, Cristiano André Da Costa, Rodrigo
gration, 19:100160, 2020. da Rosa Righi, Miromar Jose de Lima, Eduardo Silveira
[20] Patrick Luk. Design challenge of high-speed high-power den- da Trindade, and Guann Pyng Li. Predictive maintenance in
sity motor for advanced electrical submersible pump. In 2023 the industry 4.0: A systematic literature review. Computers &
IEEE International Electric Machines & Drives Conference Industrial Engineering, 150:106889, 2020.
(IEMDC), pages 1–6. IEEE, 2023.
[21] Marek Moleda, Bożena Małysiak-Mrozek, Weiping Ding,
Vaidy Sunderam, and Dariusz Mrozek. From corrective to
predictive maintenance—a review of maintenance approaches
for the power industry. Sensors, 23(13):5970, 2023.
[22] LF Pastre and A Fastovets. The evolution of esp technology
in the north sea: a reliability study based on historical data
and survival analysis. In SPE Russian Petroleum Technology
Conference. OnePetro, 2017.
[23] Marco Peixeiro. Timegpt: The first foundation model for time
series forecasting. https://towardsdatascience.com/
timegpt-the-first-foundation-model-for-time-series-forecasting-bf0a75e63b3a,
2023. Accessed: 19-02-2024.
[24] Yongyi Ran, Xin Zhou, Pengfeng Lin, Yonggang Wen, and
Ruilong Deng. A survey of predictive maintenance: Systems,
purposes and approaches. arXiv preprint arXiv:1912.07383,
2019.
[25] Luis Ribeiro and Jose Barata. Re-thinking diagnosis for future
automation systems: An analysis of current diagnostic prac-
tices and their applicability in emerging it based production
paradigms. Computers in Industry, 62(7):639–659, 2011.
[26] Simon Robatto Simard, Michel Gamache, and Philippe
Doyon-Poulin. Current practices for preventive maintenance
and expectations for predictive maintenance in east-canadian
mines. Mining, 3(1):26–53, 2023.
[27] Mehar Sahu, Rohan Gupta, Rashmi K Ambasta, and Pravir
Kumar. Artificial intelligence and machine learning in preci-
sion medicine: A paradigm shift in big data analysis. Progress
in Molecular Biology and Translational Science, 190(1):57–
100, 2022.
[28] Ahmed Salah, Ahmed Sabaa, A Samir Abd Elhaleem, Ahmed
Medhat, Ahmed Diaa, and Osama Abu-hozifa. Esp design
improvements associated with downhole slotted gravitational
filter and downhole chemical treatment prolong esps run
life in western desert of egypt: Case study. In Abu Dhabi
International Petroleum Exhibition and Conference, page
D021S051R006. SPE, 2023.
[29] Luigi Saputelli, Bravo Cesar, Nikolaou Michael, Lopez Car-
los, Ron Cramer, Mochizuki Toshi, and Giuseppe Moricca.
Best practices and lessons learned after 10 years of digital
oilfield (dof) implementations. In SPE Kuwait Oil and Gas
Show and Conference, pages SPE–167269. SPE, 2013.
[30] Luigi Saputelli, Carlos Palacios, and Cesar Bravo. Case stud-
ies involving machine learning for predictive maintenance
in oil and gas production operations. In Machine Learning
Applications in Subsurface Energy Resource Management,
pages 313–336. CRC Press, 2022.
[31] Andreas Theissler, Judith Pérez-Velázquez, Marcel Kettel-
gerdes, and Gordon Elger. Predictive maintenance enabled
by machine learning: Use cases and challenges in the au-
tomotive industry. Reliability engineering & system safety,
215:107864, 2021.

10

View publication stats