Bridging the analytics gap

Fortunately, advanced analytics are helping refiners to bridge the gap between classical first principles engineering and modern machine learning (ML) techniques, providing the means to make sense of all available data. This hybrid modelling approach empowers companies to predict equipment failures before they occur, monitor asset health in real-time, optimise throughput by reducing process variability, and make data-driven decisions that directly impact the bottom line. By applying these capabilities, refiners can achieve operational excellence that was previously unattainable with either approach by itself.

Conventional efficiency issues

Refineries face multiple operational challenges that compound to create significant efficiency losses and safety risks, and these pain points represent billions of dollars in lost opportunity across the industry.

Unplanned downtime remains the most visible and costly challenge. Critical equipment – such as heat exchangers, compressors, pumps, and control valves – degrade over time, yet traditional time-based maintenance schedules often do not capture optimal windows for intervention. For example, engineers may not detect fouling in heat exchangers until throughput constraints force rate reductions, causing millions in lost margins. Additionally, compressor failures can trigger catastrophic shutdowns, and control valve degradation can introduce process variability that cascades throughout interconnected unit lines.

Even when equipment operates within specification, excessive variation in key process parameters can reduce effective capacity, increase energy consumption, and create quality issues. Studies from the American Institute of Chemical Engineers (AIChE) indicate that reducing process variability can increase throughput by 10 - 20% without capital investment. However, identifying and addressing the root cause of variability requires analysing thousands of process variables across multiple units, a task that overwhelms traditional manual analysis methods.

Solve problems with hybrid modelling

Modern advanced analytics platforms address these challenges through an integrated approach that combines multiple capabilities: hybrid modelling that blends first principles with ML, real-time monitoring and prediction, enterprise scale deployment, and self-service analytics that empower subject matter experts (Figure 1).

Figure 1. An enterprise scale compressor health monitoring dashboard shows real-time equipment performance indicators across multiple rotating assets. Advanced analytics identify degradation patterns and anomalies, enabling engineers to detect reliability issues early and prioritise maintenance actions before failures occur.

Hybrid modelling is an evolution of the traditional binary choice between first principles and data-driven approaches. First principles models – grounded in thermodynamics, mass and energy balancing, and transport phenomena – provide transparency and physical insight. Engineers understand how these models work and trust their predictions. However, these models require detailed process knowledge, can be time-consuming to develop and maintain, and often struggle with uncertain parameters or complex nonlinear behaviour.

ML models excel at identifying patterns in large datasets and can automatically improve with more data. They handle nonlinear relationships well and can be quickly deployed across similar assets. However, they require substantial high-quality training data, their black-box nature creates trust issues, and they may fail when confronted with operating conditions outside of training ranges.

The hybrid approach leverages the strengths of both paradigms. In this approach, engineers begin with first principles equations that capture fundamental process physics. For example, when predicting heat exchanger end-of-cycle, Darcy’s Law and heat transfer equations calculate the overall heat transfer coefficient based on temperatures and flow rates. These physics-based calculations provide the structural backbone and ensure predictions remain physically meaningful (Figure 2).

Figure 2. A hybrid modelling example that combines first-principles heat-transfer calculations with predictive analytics. The calculated heat transfer coefficient is monitored over time and forecast using time-series analysis to predict when exchanger performance will reach the minimum allowable threshold, enabling proactive maintenance planning.

ML then augments the first principles model by capturing complex relationships that are difficult to model explicitly. Time-series analysis identifies degradation trends, while prediction algorithms forecast when performance will reach critical thresholds. Advanced filtering and data cleansing remove periods of abnormal operation, ensuring that models train on representative data. The result is a model that combines engineering transparency with adaptive learning capability.

In heat exchanger monitoring, for example, the hybrid approach empowers engineers to accurately predict when fouling will require cleaning. The system continuously calculates the heat transfer coefficient using first principles equations, then applies ML to model its degradation over time. When the predicted trajectory intersects the minimum allowable performance threshold, the system alerts maintenance teams with sufficient lead time to plan interventions during optimal windows. This approach has helped refineries avoid millions in lost throughput from unplanned heat exchanger constraints.

For throughput optimisation, predictive analytics help balance competing objectives. Consider a polymerisation unit plagued by fouling that forces periodic defouling procedures, each resulting in hours of off-specification production. Engineers can use historical data to model the degradation rate, calculate the optimal number of fouling-defouling cycles for a given production target, and create predictive profiles that forecast when interventions should occur. As production progresses, real-time comparison against the forecast enables dynamic adjustment of the schedule to minimise total production time and off-spec volume. Self-service analytics democratise access to these advanced capabilities. While data scientists and automation engineers play important roles, the real domain expertise resides with process engineers, reliability engineers, and operations personnel who understand the nuances of specific units and equipment.

Modern platforms provide intuitive interfaces that enable these subject matter experts to build and deploy sophisticated models without extensive programming knowledge. Furthermore, point-and-click tools for data cleansing, condition identification, and model training place analytical power directly in the hands of those closest to the problems.

Real-world results

The combination of hybrid modelling, real-time monitoring, and enterprise-scale deployment has delivered measurable results across numerous industrial operations.

Predicting viscosity to reduce product margin losses

One large chemical manufacturer operating multiple polymer production lines was facing persistent issues with product quality variability. The multi-unit process frequently produced off-specification material, but operators and engineers were unable to identify root causes using traditional analysis methods.

By using Seeq, an advanced analytics and AI platform, the team developed ML models that predicted product viscosity based on upstream process conditions. The models were trained on historical periods of on-specification production, validated against known outcomes, and deployed for real-time quality control. When the predicted viscosity deviated from target ranges, the system alerted operators to make adjustments before the material reached specification limits downstream (Figure 3).

This proactive approach reduced product margin losses by more than US$1 million/yr for a single unit producing on average 40 000 lbs/h.

Figure 3. Real-time process variability monitoring using statistical thresholds and control limits. Continuous analytics highlight deviations from expected operating ranges, helping engineers quickly identify root causes of variability and maintain stable production conditions.

Improving catalyst cycle management to increase crude processing volume

In another case, an independent refining company was struggling with catalyst bed management in its fluid catalytic cracking (FCC) unit. Uncertainty about catalyst activity and remaining useful life forced conservative operating strategies that left margins on the table.

By implementing the same advanced analytics tool, equipped with first principles calculations of weighted average bed temperature combined with predictive models trained on multiple cycle histories, the engineering team gained confidence to optimise reactor operating severity. The models automatically updated as new data accumulated, adapting to changes in feed composition and catalyst performance. Improved catalyst cycle management empowered the refinery to process additional crude volumes during periods of favourable margins, while also ensuring sufficient remaining catalyst activity during necessary turnarounds.

Scaling a predictive equipment health visualisation across the enterprise

In a final example, an international oil company with refineries spread across multiple continents deployed the same analytics platform throughout the enterprise to standardise and scale best practices. A team at one facility developed a control valve health monitoring model that identified degradation patterns before failures occurred. Using the platform’s asset hierarchy capabilities and cloud computing resources, the model was rapidly deployed to similar valves across all sites, where treemap visualisations provided management teams with at-a-glance views of valve health throughout the organisation. This helped teams prioritise where to allocate maintenance resources for maximum impact, while ensuring consistent methodology and knowledge sharing across geographical boundaries.

Combine two tools to multiply value

The convergence of first principles engineering and ML presents a transformative opportunity for refining operations where engineers are no longer forced to choose between physically meaningful models and data-driven pattern recognition. Hybrid modelling approaches deliver the best of both worlds, providing predictions grounded in fundamental process physics that adapt and improve as additional data accumulates. To achieve these results, organisations must empower their subject matter experts with self-service analytical capabilities, breaking down data silos that inhibit insight generation. With unplanned downtime costing refiners billions annually, even modest improvements in predictive maintenance deliver rapid return on investment. Process variability reduction offers a rare opportunity to increase throughput and improve margins without capital expenditure, and real-time monitoring provides proactive decision-making capabilities that predict and prevent failures to maximise uptime and profitability for refiners and other industrial operators.

Written by Shaista Mallik, Seeq.

Reference

1. ARC Advisory Group, ‘Top Technology Trends in Oil & Gas’, https://www.arcweb.com/our-expertise-industries/top-technology-trends-oil-gas

Read the article online at: https://www.hydrocarbonengineering.com/special-reports/12052026/bridging-the-analytics-gap/