A transparent Open-Box learning network provides insight to complex systems and a performance benchmark for more-opaque machine learning algorithms

Abstract

It is now commonplace to deploy neural networks and machine-learning algorithms to provide predictions derived from complex systems with multiple underlying variables. This is particularly useful where direct measurements for the key variables are limited in number and/or difficult to obtain. There are many petroleum systems that fit this description. Whereas artificial intelligence algorithms offer effective solutions to predicting these difficult-to-measure dependent variables they often fail to provide insight to the underlying systems and the relationships between the variables used to derive their predictions are obscure. To the user such systems often represent "black boxe". The novel transparent open box (TOB) learning network algorithm described here overcomes these limitations by clearly revealing its intermediate calculations and the weightings applied to its independent variables in deriving its predictions. The steps involved in building and optimizing the TOB network are described in detail. For small to mid-sized datasets the TOB network can be deployed using spreadsheet formulas and standard optimizers; for larger systems coded algorithms and customised optimizers are easy to configure. Hybrid applications combining spreadsheet benefits (e.g. Microsoft Excel Solver) with algorithm code are also effective. The TOB learning network is applied to three petroleum datasets and demonstrates both its learning capabilities and the insight to the modelled systems that it is able to provide. TOB is not proposed as a replacement for neural networks and machine learning algorithms, but as a complementary tool; it can serve as a performance benchmark for some of the less transparent algorithms.

Predicting Stability of a Decentralized Power Grid Linking Electricity Price Formulation to Grid Frequency Applying an Optimized Data-Matching Learning Network to Simulated Data

“…The steps and mathematical basic for the optimized nearestneighbour, transparent open-box (TOB) algorithm [1,43] are summarized here.…”

“…The TOB learning network is a recently introduced machinelearning algorithm successfully applied to small and mediumsized datasets [1,43]. Key benefits of the TOB algorithm are that it does not employ hidden correlations among the variables and enables forensic access to the data predictions that it makes for each data record.…”

“…The recently developed machine-learning algorithm, the transparent open box (TOB) learning network [1] is a datamatching, supervised learning algorithm that does not rely upon correlations between its input variables, prioritizing transparency and accuracy as its primary objectives. The main objectives of this study are to demonstrate the TOB's effectiveness in evaluating virtual simulations of decentralized grid performance, to predict the stability of those grids with meaningful accuracy and to provide useful insights to the relative contributions of the input variables influencing grid stability.…”

Application of mathematical and machine learning models to predict differential pressure of autonomous downhole inflow control devices

“…On the other hand, mathematical models can only be developed for one valve, and when there are several valves in a well, a separate model must be developed for each valve. In recent years, many soft computing techniques and machine-learning methods, some hybridized with efficient optimization algorithms, have been adopted as powerful approaches to predict various parameters associated with complex systems in the oil and gas industry (Rabiei et al, 2015;Jovic et al, 2016;Wood, 2018;Yavari et al, 2018;Ashfari et al, 2019;Barbosa et al, 2019;Rashid et al, 2019;Sabah et al, 2019;Yilmaz et al, 2019;Elkatatny, 2020;Gamal et al, 2020;Ghorbani et al, 2020;Mehrad et al, 2020;Moazzeni and Khamehchi, 2020;Ossai and Duru, 2020;Somehsaraei et al, 2020;Abad et al, 2021;Hazbeh et al, 2021;Mardanirad et al, 2021;Mohamadian et al, 2021).…”

Reduced-Order Modeling of Subsurface Multi-phase Flow Models Using Deep Residual Recurrent Neural Networks

“…Physics-based surrogate models are derived from high-fidelity models using approaches such as simplifying physics assumptions, using coarse grids, and/or upscaling of the model parameters (Durlofsky and Chen 2012;Frangos et al 2010;He 2013;Babaei et al 2013). Data-fit models are generated using the detailed simulation data to regress the relation between the input and the corresponding output of interest (Frangos et al 2010;Yeten et al 2005;Abdi-Khanghah et al 2018;Wood 2018). For a complete review of various surrogate modeling techniques, we refer the readers to the following papers by Asher et al (2015), Frangos et al (2010), Koziel and Leifsson (2013) and Razavi et al (2012).…”