Artificial intelligence and predictive analytics for H₂S risk management in high-hazard oil and gas facilities

Introduction

The production and processing of sour hydrocarbon feedstock entail the presence of hydrogen sulphide in the process environment – a substance with a pronounced neurotoxic action that is hazardous even upon short-term exposure. The risk depends non-linearly on concentration: at high values paralysis of the respiratory centre sets in, and the time available for protective action is reduced to tens of seconds. The problem of assessing the acute effect of sulphur compounds on the human organism under emergency conditions at natural gas fields has been discussed in the literature for several decades; contemporary reviews of hydrogen sulphide toxicology emphasise the abruptness of the transition from irritant effects to sudden loss of consciousness as concentration rises [6, p. 569-581]. A retrospective analysis of incidents shows that severe consequences are more often associated not with the absence of monitoring equipment, but with a delayed response to an exceedance that has already occurred.

The chronic character of the impact of emissions on personnel and on the population of adjacent territories is no less significant. Assessments of the influence of the petroleum refining and petrochemical industries on the ecological and hygienic state of the environment show that the system of objective monitoring of industrial emissions and discharges remains insufficiently developed, while atmospheric air pollution remains one of the principal risk factors for public health [7]. In the case of hydrogen sulphide, this uncertainty is amplified by the variability of concentrations: burst releases during purging, vessel draining or loss of flange integrity form local clouds that a stationary sensor network records only partially. Statistics of mean daily values smooth out precisely those events that determine the risk.

The classical scheme of gas monitoring is built on a threshold principle: an alarm is generated once the concentration has already reached a specified level, which makes the approach inherently reactive. Attempts to overcome this limitation have been undertaken for a long time; current work demonstrates real-time prediction of hydrogen sulphide leakage and dispersion concentrations by machine learning models trained on the results of numerical simulation and sensor observations [8]. There, the forecast is treated as an independent function of the measuring system rather than as a superstructure over the dispatcher console. Modern computational tools make it possible to return to this idea at a qualitatively different level.

Industry-wide digitalisation provides the infrastructural basis for this. Studies of the integration of digital technologies into the oil and gas sector record gains in operational efficiency, a reduction in capital and operating expenditure and the mitigation of accident risk through predictive models, while simultaneously identifying and ranking the barriers that impede their adoption – organisational, technological, financial and human-resource related [1]. The application of such models to hydrogen sulphide hazard, however, has been studied only fragmentarily: most publications are devoted to forecasting equipment failures rather than to the dynamics of the toxic gas environment. The aim of this work is to systematise approaches to constructing a system of predictive analytics for hydrogen sulphide risk at hazardous production facilities and to determine the conditions under which predictive estimates are suitable for managerial decision-making.

Materials and Methods

The study considered typical sites where hydrogen sulphide is present in process streams or is formed during operations: complex gas treatment units, oil separation and stabilisation nodes, amine treating and solvent regeneration blocks, sulphur recovery units, tank farms, flare and drainage systems, and offshore platforms handling sour products. For each group, gassing scenarios were recorded: slow accumulation in poorly ventilated pits and wells, a short-term release upon opening a vessel, and propagation of a plume from a leaking connection under light wind. This classification is of fundamental importance: the scenarios require different forecasting horizons, ranging from several minutes down to seconds.

Information was obtained from three groups of systems. The first comprises fixed gas-analytical channels with known response time constants and metrological characteristics. The second consists of personal gas detectors and wearable sensors tied to the movements of the worker and therefore characterising actual exposure rather than the state of a point in space. The third comprises contextual streams: process parameters (pressure, temperature, flow rate, valve position), meteorological observations, and the results of rounds and inspections. Experience in constructing data-driven models of hydrogen sulphide dispersion confirms that a model trained on sensor and process data reproduces the dynamics of a release far faster than direct numerical simulation, which is precisely what makes operation in real time possible [8]. The quality of the model is determined above all by the completeness and synchronisation of the data streams rather than by the complexity of the algorithm.

The methodological apparatus combined deterministic and statistical tools. Assessment of the scale of possible impact drew on a multi-criteria approach to evaluating the environmental performance of oil and gas production units, which presupposes the aggregation of heterogeneous indicators into an integral index and the delineation of the contour of the facility’s negative environmental impact [5, p. 614-625]. Transport and dispersion were described by simplified dispersion models calibrated against archived readings of the sensor network. Short-term forecasting of concentrations was performed by time series analysis methods (recurrent and gradient boosting models) using signal lags and process context. The probability of failure or drift of a measuring channel was assessed separately.

The transition from concentration forecasting to risk management required a formalised procedure. Its basis was the logic of machine learning algorithms for risk assessment in hazardous environments, which encompass the analysis of stochastic and non-stochastic hazardous factors, the evaluation of their significance and the selection of priority factors by neural network methods [3]. Hydrogen sulphide hazard was described as a combination of the probability of gassing within a zone, the duration of exposure and the severity of consequences for the worker. The results of technical diagnostics obtained by unmanned and robotic means were additionally taken into account; their application to the monitoring of hazardous facilities is described in the literature [2]. The resulting estimate was aggregated by workplace and by process operation.

Verification was carried out on retrospective samples: the model was trained on an earlier interval and tested on a subsequent one, which precludes leakage of information from the future. The metrics used were the recall and precision of exceedance detection, the false alarm rate per unit of time, and the mean lead time of the signal; the latter indicator determines whether personnel will have time to don protective equipment or leave the zone. Limitations are associated with the irregularity of archived records, differences in sensor network configurations, and the impossibility of reproducing severe emergency scenarios under field conditions, which compels reliance on model-based and simulator reconstructions.

Results and Discussion

A synthesis of the approaches makes it possible to propose a four-level architecture for hydrogen sulphide risk management. The sensor level provides measurements with health monitoring and primary filtering. The integration level synchronises heterogeneous streams, brings them to a common time grid and links them to a spatial model of the site. The analytical level contains an ensemble of models: short-term forecasting of concentrations, scenario modelling of dispersion, estimation of the probability of sensor failure, and ranking of working zones by integral risk. The managerial level converts the forecast into prescriptions – from a change in the inspection route to suspension of an operation; the gap between the third and fourth levels remains the principal obstacle to implementation.

Forecasting concentrations over a horizon of several to several tens of minutes yields the greatest effect in enclosed and poorly ventilated spaces, where the accumulation dynamics are inertial and well described by the signal history. For open sites the accuracy of point forecasting falls predictably, and it proves more productive to estimate the probability of exceeding a threshold within a given zone over a given interval rather than to predict a specific value. The probabilistic formulation is better aligned with risk-based work planning procedures and is more robust to gaps in the measurements.

The spatial incompleteness of the fixed network is partly compensated by mobile monitoring assets. The synergy of unmanned aerial vehicles, robotics and computer vision makes it possible to move from periodic inspection to continuous proactive monitoring, reducing risks to personnel and the probability of major accidents [2]. As applied to hydrogen sulphide, this addresses two tasks: the survey of zones that are themselves hazardous to enter, and the refinement of the situation where the model indicates an elevated probability of gassing. A closed loop thereby arises: the model formulates a hypothesis, the mobile platform verifies it, and the measurement result further trains the model. The effectiveness of the loop depends on the speed of platform deployment.

A separate result concerns the human factor. Predictive information does not reduce risk automatically – it must be correctly interpreted by the worker under time pressure. Multi-level industrial safety simulators are in demand here; the effectiveness of immersive virtual reality training in shaping safe behaviour during work in confined spaces has been demonstrated using the Kirkpatrick evaluation model [4]. The simulator environment makes it possible to rehearse actions in response to a probabilistic rather than a threshold signal, which is psychologically non-trivial: the worker must react to an event that has not yet occurred. Operator reactions recorded in the simulator serve as material for calibrating the admissible level of false alarms: an overly sensitive system loses its value faster than a conservative one.

The integration of predictive estimates into the occupational risk management system requires a revision of the assessment procedure itself. Machine learning based ranking of hazardous factors, proposed for hazardous production environments, makes it possible to identify priority factors and to concentrate protective resources upon them [3]. In the case of hydrogen sulphide hazard this means a transition from estimates averaged over a shop floor to estimates by operation and by workplace, taking into account the actual dynamics of the gas environment. Coupling with a multi-criteria description of the contour of the facility’s negative impact harmonises the production and environmental monitoring loops [5, p. 614-625]. In the case of hydrogen sulphide, the environmental and occupational components of risk are physically inseparable.

Implementation constraints are predominantly non-algorithmic in nature. Studies of the digital transformation of the industry point to systemic barriers alongside the positive effects of predictive models [1]. The first barrier is the quality of the source information: periods without calibration, undocumented sensor replacements, desynchronisation of archives. The second is interpretability: a decision to suspend work cannot rest on an opaque model output; an explanation in terms of measurable quantities is required. The third is regulatory: a predictive estimate does not possess a status comparable to that of a reading from a verified measuring instrument and is used as auxiliary information. The fourth relates to ecological and hygienic regulation, where the insufficient development of a system of objective emission monitoring limits the verification of models against independent measurements [7]. The removal of these constraints is a precondition for the transition to routine operation.

Conclusion

The analysis shows that predictive analytics, as applied to hydrogen sulphide hazard, addresses a different problem from that of traditional gas monitoring. Threshold alarming registers an event that has already taken place, whereas a predictive model estimates the probability of its occurrence and leaves a reserve of time for protective action. The greatest return is achieved in enclosed spaces with inertial accumulation dynamics, while for open sites a probabilistic formulation of the problem is preferable. The proposed four-level architecture links the sensor, integration, analytical and managerial loops, the weak link remaining the conversion of a forecast into a regulated action: without formalising this transition, the accuracy of the model is not converted into a reduction of risk.

The second conclusion concerns the complementarity of the solutions. Neither the fixed network, nor mobile robotic assets, nor simulator-based personnel training provides the required level of protection in isolation: the result is delivered by their coupling into a feedback loop in which the forecast initiates verification, verification refines the model, and the simulator builds readiness to act upon an anticipatory signal. Occupational risk assessment must at the same time be detailed down to the level of operations and workplaces, which is consistent with neural network methods of ranking hazardous factors.

Further research should be concentrated in three directions: validation methodologies for predictive models of the gas environment that are suitable for regulatory recognition; interpretable models whose output is formulated in measurable terms; and assessment of the economic efficiency of predictive systems with allowance for prevented damage. The admissible proportion of false alarms, which determines personnel trust in the system and ultimately its operability, warrants separate attention.