Source-linked AI summary
Improving Energy Efficiency of Oil Platforms Through Optimal Loading of Diesel Generators Using Machine Learning and Search Algorithms
Khivishta Boodhoo, Josh Plumbly, Nicholas Watson
TL;DR
Offshore oil-platform research has focused more on maximising production than reducing platform diesel consumption. This study uses ML models and search algorithms to optimise diesel-generator loading, reporting an approximate 27% fuel-consumption reduction while noting limitations from a limited, not fully representative dataset.
Problem
Reducing energy use by offshore oil platforms has received less attention than maximising oil and gas production, despite diesel generators supplying platform electricity.
Method
The study combines a tailored ML methodology for diesel-consumption prediction with search-generated daily diesel-generator load combinations and visualisation.
Results
An approximate 27% reduction in fuel consumption was achieved for the evaluated daily load combinations.
Takeaways & Limitations
The findings indicate an opportunity to improve offshore-platform energy efficiency by optimising diesel-generator loading.
Takeaways & Limitations
The study relies on a limited dataset that may not be fully representative, and more actual recorded data could enhance model performance.
Abstract
from arXiv · showhide
Rising energy demand, fossil fuel depletion and climate change highlight the need for more efficient energy production and consumption. Offshore oil and gas platforms face challenges related to inefficient energy use, system failures, accessibility and environmental impact. Machine learning (ML) offers opportunities to improve the safety, sustainability and efficiency of these systems; however, previous research has largely focused on increasing oil production rather than reducing energy consumption on platforms. This study investigates the use of ML and search algorithms to improve diesel efficiency on an offshore oil platform. Data collected over 18 months from a platform in Scotland were analysed, focusing on four diesel generators as the primary diesel-consuming equipment. Following exploratory data analysis and outlier detection, regression models were developed to predict daily diesel consumption for different generator power loads. Multiple Linear Regression and Artificial Neural Networks achieved the best predictive performance compared with Extra Trees Regression, Extreme Gradient Boosting and Random Forest. Search algorithms were then used to identify combinations of generator power loads that minimised daily diesel consumption. The results showed an average diesel saving of 27% per day compared with the worst daily power-load combinations, equivalent to approximately 24,000 litres/day. These findings demonstrate significant opportunities for improving energy efficiency on offshore oil platforms using ML-based optimisation.
1. Introduction and Background
Prior research has applied data-driven and optimisation methods to fuel use, but offshore-platform studies have largely prioritised oil production rather than directly minimising diesel consumption. This study addresses that gap with ML and search algorithms for optimising diesel-generator loading.
- Fuel-consumption prediction studies in vehicles, buses, aircraft, ships, and marine diesel engines demonstrate the usefulness of data-driven models but have limited transferability.
- Offshore oil platforms require substantial energy, and diesel generators provide electricity while efficient operation has received limited consideration.
- Previous oil-platform research has mainly focused on production, platform design, process simulation, energy systems, and power-network stability.
- Prior optimisation approaches include Improved Harmony Search and Genetic Algorithms, but their robustness can be constrained by initial-population sensitivity and premature convergence.
- Methods directly optimising diesel consumption through daily generator-load distribution on oil platforms remain insufficiently studied.
- The proposed methodology combines ML and search algorithms with a visualisation tool to identify optimal generator-load distributions for minimal diesel consumption.
2. Methodology
The methodology section introduces the studied oil platform, maps diesel flow to equipment, describes the datasets, and outlines the modelling strategies.
- The study first presents the oil platform and a process-flow diagram showing diesel flow to different equipment.
- It then describes the relevant datasets before outlining the modelling strategies used.
2.1. Overview of Oil Platform and Dataset
The study examines an FPSO platform where diesel is mainly consumed by diesel generators, with additional use by process heaters and inert-gas generators. The dataset combines generator-load records with manually recorded daily diesel use and equipment-status features.
- The selected FPSO uses five Wartsila diesel generators, but broken generator DG-C was excluded, leaving four generators for modelling.
- Diesel is mainly consumed by the diesel generators, while bow thrusters and fire-water pumps use negligible amounts.
- Inert Gas Generators burn diesel to create a low-oxygen blanket during cargo offloading, with a flow rate of 10,392 kg/day.
- Equipment-status features, including oxygen concentration near the IGGs and PFH exhaust temperatures, supported daily diesel-consumption calculation.
- Generator power loads were recorded at irregular intervals from 2010, while overall daily diesel usage was manually recorded from 2021.
2.2. Method Designed to Reduce Diesel Consumption on the Oil Platform
The study preprocesses platform data, trains regression and ML models to predict diesel consumption, and applies search-generated generator-load combinations to identify low-consumption operating points. Results are visualised with a multidimensional plot.
- The workflow calculates daily generator loads and diesel consumption for Process Fired Heaters and Inert Gas Generators before preprocessing.
- Search algorithms generate daily generator-load combinations, which are passed to the best-performing ML models for diesel-consumption prediction.
- A Parallel Coordinate Plot visualises predicted consumption and the associated generator-load combinations.
- Preprocessing includes visualisation, correlation analysis, feature engineering, and Mahalanobis-Distance outlier removal.
- Multiple Linear Regression, XGBoost, Random Forest, Extra Trees Regressors, and Artificial Neural Networks predict diesel consumption.
- The workflow identifies the least, most, and median predicted diesel consumption with their corresponding load combinations.
2.2.1. Data Preprocessing
Data preprocessing aligned daily generator-load features with daily diesel consumption, removed weak or redundant features and anomalous observations, and assessed feature relationships before modelling. After preprocessing, total daily loads showed a strong correlation with diesel consumption, while seasonality produced consistent model predictions with or without the feature.
- Feature Engineering: Daily generator measurements recorded at varying intervals were converted into average daily load values representing energy used or required.The conversion used the relationship E(kWh) = P(kW) × T(h), where E is daily energy, P is generator power load, and T is operating time.
- Feature Engineering: Approximately 59 m3/day of constant PFH diesel consumption had no useful relationship with the target and was excluded from subsequent modelling.Its constancy meant it did not provide useful variation for predicting daily recorded diesel consumption.
- Outlier Detection and Removal: Mahalanobis Distance was used to identify and remove outliers after visual checks found inconsistent relationships between stacked generator loads and diesel consumption.Potential sensor malfunctions or recording errors were indicated when loads and recorded consumption moved inconsistently, including zero-load or near-zero-consumption cases.
- Feature Engineering: Seasonality was tested as an additional feature, and predictions were consistent with or without it, indicating minimal weather influence on diesel-consumption predictions.The best-performing model was trained, tested, and validated in both conditions.
2.2.2. Machine Learning Modelling
The study evaluates five machine-learning models for regression, combining a linear model, tree-based ensembles, and an artificial neural network. Models were trained and validated using data splits and cross-validation, with model selection guided by predictive performance.
- Model selection: Five ML models spanning statistical, tree-based ensemble, and neural-network approaches were explored for regression.The models included Multiple Linear Regression, XGBoost, Random Forest, Extra Trees Regressors, and an ANN.
- Tree-based ensembles: XGBoost, Random Forest, and Extra Trees Regressors were selected because tree ensembles can identify patterns without requiring an initial relationship assumption.The models aggregate decision-tree predictions and were considered robust to some randomness or error in the data.
- Artificial Neural Network: The ANN was included because adjustable neuron weights allow it to learn intricate relationships between inputs and outputs.During training, the weights are iteratively adjusted to reduce differences between predictions and actual outcomes.
- Training and validation: The dataset was split into 0.7 training and 0.3 testing subsets, with validation and cross-validation used during model assessment and grid search.A fixed random seed supported reproducible splits, while cross-validation helped assess hyperparameter combinations.
- Multiple Linear Regression: Multiple Linear Regression assumes linear relationships between independent variables and the dependent variable.Its general form combines an intercept, feature coefficients, input variables, and an error term.
2.2.3. Deployment
The deployment procedure generates feasible load combinations for four diesel generators, predicts their diesel consumption with the best-performing ML model, and searches for efficient loadings at equal total demand. Best, median, and worst combinations are then compared to estimate potential savings.
- Deployment inputs: The deployment model predicts daily diesel consumption for combinations of loads on DGs A, B, D, and E.The inputs represent different load distributions among the four generators.
- Input generation: Load values were generated across each generator’s observed minimum-to-maximum range using 3% intervals and Cartesian products.The interval was chosen partly because computational resources limited the number of generated values.
- Feasibility filtering: Generated combinations were filtered to the original dataset’s total-load thresholds of 68,000 kW to 400,000 kW.This produced 1,230,288 feasible daily load combinations.
- Consumption prediction: Each feasible combination was normalised, passed through the best-performing ML model, and denormalised to predicted diesel consumption in m3.The model therefore maps generator-load combinations to estimated actual diesel usage.
- Search optimisation: Brute-force search identifies the least-consuming load combination among combinations with the same daily total load.Because generators differ in efficiency, equal total loads can produce different predicted diesel consumption.
- Visualisation and savings: Best, median, and worst load combinations are visualised in a Parallel Coordinate Plot and used to calculate average daily fuel savings.Savings are computed by subtracting best-case predicted consumption from worst-case predicted consumption at the same total load.
3. Results and Discussion
The results identify ANN and MLR as the strongest predictive models, with ANN selected for deployment despite its greater training time. Applying the model to generator-load combinations reveals substantial potential fuel savings, while comparisons across studies remain difficult because datasets and problem scopes differ.
- Model performance: MLR and ANN performed better than XGBoost, RF, and ETRs on the original dataset.The study therefore focused deployment decisions on these two models.
- Model performance: ANN achieved an R2 score of 0.81 and an MAE of 0.07, compared with MLR’s R2 score of 0.80 and MAE of 0.08.ANN also had the lowest MSE, RMSE, and MAE and the highest R2 among the studied algorithms.
- Model performance: ANN performance indicated stronger ability to capture non-linear relationships, whereas RF had the worst reported performance among the studied models.RF had the highest MSE, RMSE, and MAE together with the lowest R2 score.
- Comparison with prior work: Comparisons with other studies are challenging and unreliable because performance depends on datasets, model complexity, and problem scope.The authors therefore note that the best-performing model varies according to circumstances.
- Load optimisation: Equal total daily loads can yield different predicted diesel consumption because the generators differ in efficiency.The efficient combination is therefore determined by how load is distributed across the generators, not only by total demand.
- Fuel savings: Fuel savings peaked at approximately 250 MW of daily total power before declining.The reported relationship between daily total diesel consumption and required daily loads was otherwise described as linear.
- Fuel savings: 9 m3/day, or 9000 litres/day, represented approximately 12% savings versus average real-world diesel consumption.The real-world dataset averaged approximately 74 m3/day, while the best-case scenario averaged about 65 m3/day.
4. Conclusion
This study applies machine learning and exhaustive search to identify diesel-generator loading strategies that reduce consumption on an offshore Scottish platform. The results indicate substantial savings, while limited data and platform-specific conditions constrain generalisability.
- Method: ML models predicted diesel consumption for different generator power loads, with Multiple Linear Regression and Artificial Neural Networks performing best among the models explored.The comparisons included Random Forest, XGBoost and Extra Trees Regression.
- Optimisation: Brute-force exhaustive search identified generator power-load combinations that minimised predicted daily diesel consumption.The search evaluated diverse load combinations generated from the best-performing ML model.
- Results: 21% reduction was achievable when average best load combinations replaced average median load combinations.This comparison used combinations consuming the least amount of daily diesel versus median load combinations.
- Results: 27% reduction in fuel consumption was achieved by selecting the best load combinations instead of the worst combinations.The best combinations consumed the least daily diesel, while the worst consumed the most.
- Application: An interactive visualisation tool was developed to show diesel-use configurations for the least, medium and greatest daily consumption.The tool was intended to support practical assessment of generator power-load configurations and anticipated daily total loads.
- Limitations: The methodology relies on a limited, not fully representative dataset from an FPSO platform in Scotland, so robustness and applicability to platforms with different generators require further investigation.The authors identify substantial datasets and diverse oil platforms as priorities for future work.
5. Data Availability
The article’s datasets belong to a specific oil platform in Scotland.
- The datasets related to this article belong to a specific oil platform in Scotland.
6. Declaration of Competing Interest
The authors report no known competing financial interests or personal relationships that could have influenced the reported work.
- The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.