Ildar Saiapov
Leader Specialist, Student
Russneft, Skoltech
I am currently actively working on a project at Russneft to rank wells for geological and technical measures (GTM).
I have experience in ML and DL: knowledge of basic machine learning algorithms and libraries.
I have well-developed communication and organizational skills. I am fluent in English, both written and spoken.
Graduated Moscow Gubkin University wih a honour (bachelor, Petroleum Engineering), now I study at Skoltech, Petroleum Engineering as well, MSc2
Also I balance my study wih a job at the Russneft company as a lead specialist
Participates in
TECHNICAL PROGRAMME | Energy Technologies
Smart Infrastructure for the Future Energy Industry: Digitalisation & Innovation
Forum 18 | Digital Poster Plaza 4
27
April
15:30
17:30
UTC+3
The goal of my project was to replace manual calculations for each new time interval of new well operations calculations with a machine learning model prediction that would provide recommendations on the best wells at each company site for each new date.
objectives:
The models were tested on data from multiple fields with significantly different geological properties. Data on more than 3,000 oil well operation events were collected and aggregated, including 1,739 were BZT (bottomhole zone treatment, in particular acid treatment), and the rest were measures to intensify oil production (IOP) by reducing bottomhole pressure. For each oil well operation, historical data on production and injection, parameters of the measures taken, and reservoir characteristics were collected and prepared. The target feature (additional production) was cleaned of outliers using the properties of quantile distribution. Unlike methods based on normal distribution, this approach is not sensitive to absolute values and the scale of outliers, which allows for more effective elimination of anomalies.
Another key task was to take into account the mutual influence of wells. Traditional correlation calculation methods required significant computing resources to process the entire history of wells. Therefore, an original algorithm based on fast Fourier transform (FFT) was developed to calculate the optimal radius of influence of production and injection wells and to estimate the correlation between them. All calculations were transferred to graphics processing units (GPUs), which accelerated the calculations by more than 10–20 times.
Various machine learning methods were tested to predict the success of GTM: gradient boosting (CatBoost) for regression and classification tasks, as well as ranking models (Learning to Rank).
It has been shown that direct prediction of the value of additional production (regression) results in high error (RMSE, MAE), which is due to the complexity of the physics of the process and the limited data available. The classification model predicting the probability of events success demonstrates acceptable accuracy (Accuracy ~0.7–0.8), but has a significant spread in predictions. The best results are achieved using a recommendation ranking model, which does not predict exact values, but ranks wells in descending order of expected relative efficiency (additional production at the field). This approach allows selecting a limited number of the most promising candidates for GTM. This recommendation model is distinguished by its ability to compare wells with each other (it calculates an individual loss function that takes into account the rank of the well - Learning to Rank).
objectives:
- formulate a specific task to be solved and collect the necessary data from databases.
- conduct data analysis and train machine learning models to predict well operations success.
- develop an algorithm for accounting for the influence of neighboring wells on the target well.
- automate data collection and model forecasting
The models were tested on data from multiple fields with significantly different geological properties. Data on more than 3,000 oil well operation events were collected and aggregated, including 1,739 were BZT (bottomhole zone treatment, in particular acid treatment), and the rest were measures to intensify oil production (IOP) by reducing bottomhole pressure. For each oil well operation, historical data on production and injection, parameters of the measures taken, and reservoir characteristics were collected and prepared. The target feature (additional production) was cleaned of outliers using the properties of quantile distribution. Unlike methods based on normal distribution, this approach is not sensitive to absolute values and the scale of outliers, which allows for more effective elimination of anomalies.
Another key task was to take into account the mutual influence of wells. Traditional correlation calculation methods required significant computing resources to process the entire history of wells. Therefore, an original algorithm based on fast Fourier transform (FFT) was developed to calculate the optimal radius of influence of production and injection wells and to estimate the correlation between them. All calculations were transferred to graphics processing units (GPUs), which accelerated the calculations by more than 10–20 times.
Various machine learning methods were tested to predict the success of GTM: gradient boosting (CatBoost) for regression and classification tasks, as well as ranking models (Learning to Rank).
It has been shown that direct prediction of the value of additional production (regression) results in high error (RMSE, MAE), which is due to the complexity of the physics of the process and the limited data available. The classification model predicting the probability of events success demonstrates acceptable accuracy (Accuracy ~0.7–0.8), but has a significant spread in predictions. The best results are achieved using a recommendation ranking model, which does not predict exact values, but ranks wells in descending order of expected relative efficiency (additional production at the field). This approach allows selecting a limited number of the most promising candidates for GTM. This recommendation model is distinguished by its ability to compare wells with each other (it calculates an individual loss function that takes into account the rank of the well - Learning to Rank).
