Predicting Environmental Impact of Hazardous Liquid Pipeline Accidents: Application of Intelligent Systems

In the United States (US), the hazardous liquid pipeline network has grown by 15% in miles and 30% by operator count since 2010 (PHMSA 2018a). The Pipeline and Hazardous Material Safety Administration (PHMSA) regulates approximately $3.47\times {10}^6\mathrm{km}$ of hazardous liquid pipelines and a total of 529 hazardous liquid pipeline operators.

Pipelines have the lowest failure rates among fuel transportation methods, on the order of ${10}^{-3}$ and ${10}^{-4}$ accidents per km per year for onshore and offshore pipelines, respectively (Belvederesi 2017; Green and Jackson 2015; Singleton 2013). However, the receiving environments, communities, and oil companies can be significantly impacted by a single accident (Belvederesi et al. 2018; TransCanada 2017; PHMSA 2018b; EPA 2015). The objective of this study is to predict environmental consequences due to hazardous liquid pipeline accidents based on pipeline explanatory variables, such as commodity transported, accident location, nominal diameter, nominal wall thickness, and material-release method. These explanatory variables were selected following a thorough statistical analysis of the relationships between pipeline design and failure (Belvederesi et al. 2018), which informs on the data availability and significance. Moreover, this set of explanatory variables represents the basic pipeline design and the information most commonly gathered by regulatory agencies after pipeline accidents worldwide (Belvederesi 2017), which helps model reproducibility in other countries. An adaptive neuro-fuzzy inference system (ANFIS) was developed to predict water and soil contamination and adverse impact on wildlife based on the aforementioned explanatory variables. Conventional approaches to system modeling, which include probabilistic analysis, are often unsuitable for complex and uncertain systems, especially in environmental analysis (Suparta and Alhasa 2016; PreventionWeb 2018). By using a fuzzy inference system (FIS) and artificial neural networks (ANN) together, it is possible to control and overcome the complexity and uncertainty of these systems (Jang 1991, 1993). However, when using FIS, there are no mathematical methods to transform the experience and knowledge of human experts into the required if–then rule-based approach, and the lack of adaptability of the learning algorithm in tuning membership functions to minimize the error often limits the applicability of FIS. To address these limitations, ANFIS was employed in this study. ANFIS is classified as artificial intelligence (AI) because its capability for processing nonlinear and complex information is similar to the human brain. ANFIS uses fuzzy conditional statements (if–then rules) to capture the imprecise modes of reasoning behind the human ability to make decisions in an environment governed by uncertainty and imprecision. The set of if–then rules is based on data, which represent human knowledge derived from past experience. Membership functions must be developed from the available data on which the model is based (fuzzification). While membership functions used in FIS are generated by human expertise, membership functions for ANFIS are generated by adaptive neural networks (NN). In addition to ranges, ANFIS also generates probability distribution types and parameters.

The literature offers some examples where ANFIS or similar approaches were adopted as intelligent systems in different fields of application. Alizadeh et al. (2018) demonstrated how machine learning can assist environmental management and monitoring tools by providing accurate predictions for water quality parameters. Wijayasekara and Manic (2014) studied how to develop membership functions using classical statistical methods. They investigated data set coverage, complementarity, and relative dissymmetry, and used this information as metrics for understandability in the generated membership functions (Wijayasekara and Manic 2014). Liu and Zhang (2015) developed an ANFIS model-based data-driven approach to control the welding process. They concluded that, compared with linear models, to control welding processes—welding speed and weld pool characteristics parameters—an iterative local ANFIS model using k-means clustering provides better results for modeling performance (Liu and Zhang 2015). Sharkawy et al. (2014) compared three algorithms to predict the surface roughness of end-milling process: a radial basis function neural network (RBFN), ANFIS, and a genetically evolved fuzzy inference system (G-FIS). Among these techniques, they found that RBFN is the most successful at performing surface roughness assessment (Sharkawy et al. 2014). Dastorani et al. (2010) studied how to predict missing flow data of gauging stations using data from nearby hydrometric stations to train an ANN and an ANFIS model. In this study, the researchers concluded that ANFIS is superior for estimating missing data, although ANN produces results with a high level of accuracy (Dastorani et al. 2010). Badde et al. (2015) compared FIS and ANFIS to predict compressive strength of ready-mix concrete (RMC). ANFIS with Gaussian membership functions could predict the 28-day compression strength of ready-mix concrete with satisfactory performance, according to the authors (Badde et al. 2015). ANFIS was also applied in the oil and gas industry to assess and optimize pipeline system performance, predict corrosion rates of steel pipelines, and reduce energy consumption for gas transportation. For example, Baghmolaei et al. (2014) developed both FIS and ANFIS models to minimize fuel consumption of gas turbines, and compared these two techniques with ANN. The results of modeling by ANN, FIS, and ANFIS showed that ANN with the genetic algorithm code has the best performance (Baghmolaei et al. 2014). He et al. (2012) studied how RBF and ANFIS could predict the corrosion rate of underground steel pipelines. Results showed that ANFIS can more accurately predict corrosion rate considering changing corrosion factors (He et al. 2012). Baghban et al. (2017) proposed a study that aimed to develop an ANFIS-based model to predict the breakthrough curves for rhamnolipid adsorption over activated carbon. The models’ accuracy (absolute average deviation of 1.98%) shows that ANFIS would be a more reliable method to predict breakthrough curves compared with ANN and group method data handling (GMDH). Shahrak et al. (2018) applied ANFIS to model water vapor uptakes in different porous metal–organic framework materials, showing a high coefficient of determination ($R^2=0.90$). A multilevel adaptive neuro-fuzzy inference system was developed by Nabavi-Pelesaraei et al. (2019) to predict various environmental, energy, and economic indices of large-scale food production systems, showing high accuracy with coefficients of determination between 0.91 and 0.98.

Predicting environmental consequences due to hazardous pipeline failures can be challenging. To develop a predictive model, it is necessary to have a robust, complete, and informative data set containing information gathered from past accidents. Belvederesi et al. (2017, 2018) investigated the limitations that derive from inadequate data and their implications for accident response actions, and conducted a thorough statistical analysis of trends in hazardous liquid pipeline accidents in relation to the pipeline design characteristics. Major findings led to the development of the model presented in this study.

In this section, results are reported for each output type: soil contamination, water contamination, and impact on wildlife. Table 2 summarizes AIP and AVP for each model. Similarly, Table 3 offers a detailed review of ANFIS performance in predicting whether the environmental contamination occurred according to actual data. These results are discussed in the next section.

Linguistic variables (biofuels, crude oil, HVL, refined petroleum products, carbon dioxide, aboveground, tank, transition area, underground, leak, mechanical puncture, other, rupture, and overfill or overflow) are converted to categorical variables for calculation purposes. The following legend of these variable names helps in interpreting the results reported in this section:

Tables 4–6 provide the testing data set used to verify the model’s performance for each analysis in the sections below. Each line contained in Tables 4–6 represents a hazardous liquid pipeline failure. Each line, or failure, provides commodity transported, location where the accident occurred, diameter-to-wall-thickness ratio, operating pressure, and mode with which the material was released. The predicted output is compared with the actual output and reported as a green tick in case of matching outcomes and a red X in case of mismatch (1 stands for Yes and 2 stands for No).

A predictive model was developed to evaluate whether an environmental impact is likely to occur in case of hazardous liquid pipeline failures. Explanatory variables, including commodity transported, accident location, diameter-to-wall-thickness ratio, maximum operating pressure, and material release mode were used to train and validate an ANFIS model. Three different outcomes were investigated: soil and water contamination and adverse effects on wildlife.

Results show that the model is capable of accurately predicting outcomes for the three impact types, as follows:

Data used to test the model show that approximately 65% of hazardous liquid pipeline accidents resulted in soil contamination, 32% resulted in water contamination, and 15% resulted in adverse impacts on wildlife. In addition, 57% of incorrectly predicted outcomes for soil contamination occurred on pipelines transporting crude oil and located underground. This combination of the two parameter values is also the most recurrent in the training database (52% of accidents). Similarly, the results of the model for soil contamination show that 79% of incorrectly predicted cases include the combination of being located underground and with leak as the release mode. These instances are the most commonly recurring combinations of this pair of parameters in the training data set (73% of the training data set). This means that the model has a poor capacity to predict rare outcomes because, when there is a set of explanatory variables that rarely leads to an outcome, the model cannot make an accurate prediction for that rare event. This limitation could be further investigated by focusing the analysis on those sets of explanatory variables that rarely lead to soil contamination. The results for the water contamination model do not show any particular trend or commonalities similar to the soil contamination and wildlife impact models. In fact, the sets of variables that lead the model to the incorrect outcome are always diverse. Finally, the low number of accidents in the training data set that had adverse effects on wildlife (4.8%) could explain why the model for predicting adverse impacts on wildlife was 100% accurate. This should be further investigated by using a different set of accident data to test the model, although PHMSA reports few cases of impact on wildlife during the past 10 years.

The selected set of explanatory variables used in this study represents the optimal combination to obtain the most accurate results. However, accident location and release mode are less informative variables than others because more than 80% of accidents involved pipelines that failed underground, and more than 50% of the pipelines failed by leaking. The low variance shown by these variables would lead to the assumption that they do not provide substantial information to the model, although, by including them in the analysis, the model performs more accurately.

Predicting adverse environmental impacts of hazardous pipeline accidents is challenging in the absence of an adequate, complete, and informative data set that gathers information from past accidents. Because the model presented in this study is data driven, a major limitation is presented in the training/validating data set; missing data and incorrect reports introduce conflicting scenarios into the model that, as a consequence, it is not able to accurately predict, especially in the case of rare events. Currently, comprehensive design information must be collected by regulators after pipeline accidents, including pipeline material; despite this, a large number of missing data points makes it challenging to include certain design variables as inputs for the model. Belvederesi et al. (2018) show that, over time, regulators are becoming stricter with regard to reporting and the quality and quantity of information collected after an accident. For this reason, future iterations of this model should include more explanatory variables as they become available, such as pipeline material, and the model should be updated with newly collected data to ensure that predictions will be accurate in addition to the model being computationally efficient.

During pipeline planning and design, this model could provide an alternative method to predict the hazards and, consequently, the level of risk that a pipeline would pose to the surrounding environment. In those instances where the failure of hazardous liquid pipelines would be likely to lead to adverse environmental outcomes, additional protective actions should be adopted for both planned and existing pipelines. Protective measures can include installing casings on water-crossing pipelines and improving or reinforcing leak detection systems. Additionally, in the case of planned pipelines, an alternative combination of design characteristics should be evaluated, especially in environmentally sensitive areas.

All data, models, or code generated or used during the study are available in a PHMSA repository online in accordance with funder data retention policies. They are available at https://www.phmsa.dot.gov/data-and-statistics/pipeline/gas-distribution-gas-gathering-gas-transmission-hazardous-liquids.

The authors acknowledge the contribution of Dr. Petr E. Komers, president of MSES, Inc., for his ongoing emotional, professional, and financial support.