Development of a reforming plant optimization model based on deep neural networks

Goal

Optimizing the plant's operational process to minimize gas consumption

Results

Reduction in gas consumption by 1.35%

Development of a digital twin of the reforming process, which enabled computer optimization of the process by changing control parameters

Challenge

The subject of this paper is a model for optimizing a reforming plant based on deep neural networks.

The amounts of equivalent gas supplied in the reforming process were too high in relation to expectations for maintaining production rates within appropriate ranges, including in particular the goals of maximizing the amount of hydrogen-nitrogen mixture produced. The goal of the project was to optimize the plant's operating process in such a way as to minimize gas consumption.

Solution

A digital twin of the reforming process was developed, which enabled computer optimization of the process by changing control parameters. The digital twin divided the reforming process into stages. Each of these stages had input, output and control variables appropriately modeled by artificial intelligence algorithms. Based on the values of the input variables, it returned optimal control values for a defined cost function for which the value is close to the minimum.

About ReliaSol ~ Azoty Tarnów Group

“Reliability Solutions has been working with Azoty Group in Tarnow for more than two years. The cooperation has primarily involved complex data analysis-applications to solve key technical machine maintenance problems, including minimizing the number of breakdowns, improving machine availability and maximizing production efficiency. Cooperation with Reliability Solutions has always gone smoothly, and the services offered have been of high quality.”

1.1 SOLUTION STRUCTURE

The implemented approach is to model each output variable separately using all the information available in a given block (i.e., the values of control variables, input variables from previous blocks and the current block).

Validation of the models created in this way was performed based on the part of the data that was not used in any way in their learning process. Thanks to this approach, it was possible to determine with high probability the degree of generalization of the model and, consequently, its predictive ability.

In the next step, the blocks modeled in this way were combined into the structure shown in the diagram below. This approach allows the prediction of all the values of the output variables for which the models were created using only the necessary values of the input and control variables.

This model could then be used in tandem with the optimization module to find the optimal control of the reforming process according to the penalty function presented in the next section.

The final step is to implement the system alongside the running process. In this case, the optimization algorithm will continuously read the values of the input data and, based on this, return the optimal process control in real time.

1.2 STAGES OF PROJECT IMPLEMENTATION

1. Data preprocessing - delays. Lag analysis techniques: (correlation analysis, regression using decision trees, linear lasso regression)

2. Metamodeling - error measurement: Neural networks (MLP networks, convolutional networks, recurrent networks (including LSTM), deep convolutional networks (with residual connections)), Forward selection.

3 Optimization - results. The adopted metamodeling methods made it possible to obtain knowledge of the process flow, while the prepared optimization algorithm developed a solution that reduced gas consumption by 1.35%.

BENEFITS

In addition to technical sophistication, our solutions focus on the economic benefits of cost reduction:
- 80% minimization of failures
- 55% minimization of downtime
- 25% increased production profits
- >10X ROI
- Prediction of potential failures of generating units, as well as the power transmission and distribution system,
- identification of failure causes (elimination of the root cause of failure),
- multi-criteria and holistic control optimization - maintenance of optimal production levels, minimization of pollution, flexible response to emerging changes, forecasting of demand for energy and/or utilities (steam, gas, water)

About company

ReliaSol is a company established in response to the growing need to increase the efficiency of machines and installations in industry.

We provide software and services that accelerate the digital transformation process. The applications we create combine real data from machines, sensors, event reports and automate drawing conclusions. The result of their work is data visualization, event prediction and monitoring of the optimal operating range of machines and entire production lines.

We will be happy to answer your questions

Reliability Solutions Sp. z o. o.
ul. Królewska 57, 30-081 Kraków, Polska