Συντάχθηκε 29-09-2021 14:56
Τόπος: Εξ’ αποστάσεως - Ανοιχτή στο κοινό
Σύνδεσμος τηλεδιάσκεψης
Έναρξη: 07/10/2021 14:00
Λήξη: 07/10/2021 15:00
Title: Artificial Intelligence for downhole tools
Τίτλος: Τεχνολογίες τεχνητής νοημοσύνης στην ανάπτυξη εργαλείων εντός της γεώτρησης για τον έλεγχο παραγωγής σε ταμιευτήρες πετρελαίου / φυσικού αερίου.
Εξεταστική Επιτροπή:
Γιώτης Ανδρέας, Επ. Καθηγητής Σχολής Μηχανικών Ορυκτών Πόρων Π.Κ. (επιβλέπων)
Γαγάνης Βασίλης, Επ. Καθηγητής Σχολής Μηχανικών Μεταλλείων-Μεταλλουργών Ε.Μ.Π.
Μαρινάκης Δημήτρης, ΕΔΙΠ Μηχανικών Ορυκτών Πόρων Π.Κ.
Abstract:
The introduction of revolutionizing methods in the Gas and Oil sector is a challenging task, especially when these methods interfere with high-risk activities in terms of operation, investment, personnel welfare and the environment. Despite the great caution in introducing new technologies, the oil and gas industry remains open to new technologies to optimize and streamline existing processes. The low price period for oil and gas that came along with the Covid-19 restrictions further fueled the rate at which the specific industry adopts new technologies in order to cope with such challenges.
The scope of the thesis is to review the artificial intelligence methods that apply to downhole equipment in the production sector of the oil/gas industry. In the first part of this work, the history of the two distinct disciplines (oil and gas and the artificial intelligence) will be connected throughout a timeline and explain the one can benefit from the other. Subsequently, two case studies will be discussed in order to study the efficiency of AI in production and workover.
The first case study is a virtual flow meter, which is a convenient substitute of physical sensors it uses available data in instrumented wells to predict other measurements (oil, water and gas flowrates). Most of the oil and gas production flowrates change with the production time, so it is necessary to recalibrate the virtual flow meter model periodically. The model runs on a dataset generated from a physics-based software (Pipesim software). The dataset was initially modified in python to make the dataset more suitable for introducing it to the virtual flowmeter model in order to set the initial parameters of the instrument. The same python code can be implemented to predict the behavior of an ESP throughout the lifetime of a field and to optimize the performance of an ESP according to the current reservoir and well conditions. The virtual flowmeter aims at predicting the multiphase flowrates based on the readings of the ESP pump downhole. Several models were considered as candidates for the virtual flowmeter. The final model (MLP) was then selected after cross-validation and hyperparameters tuning of each candidate model. The accuracy of the selected model exceeded 99% of the dataset values.
The second case study is a Bayesian network model for predicting the root cause of ESP breakdown (or stoppage). The results match with the actual assessment of the service engineer, when the latter inspected the motor caused after the pump’s failure. A third case study describes a reinforcement learning technique to autonomously make decisions for an injection well in order to keep the reservoir pressure within an optimal range, while decrease the cost. The model was able to reduce the cost by 57%.
