Prof Yong Yan, University of Kent, shares his report blog as an output from his project report for Monitoring of CO2 flow under CCS conditions through multi-modal sensing and machine learning, which was funded under our Flexible Funding 2020 call:
“Following the approval of my application for the UKCCSRC Flexible Funding 2020, I conducted the proposed work in collaboration with North China Electric Power University (NCEPU) and the industrial partners from 1 May 2020 to 31 August 2021. The following is a report on the activities that were undertaken during the project. Accurate measurement and real-time monitoring of CO2 flow across the entire CCS chain are essential to gain an in-depth understanding of the physical and chemical characteristics of CO2 flow in pipelines. Advances in these areas will also ensure the success of large-scale injection projects and enable the long-term monitoring of captured CO2 in storage sites. However, due to the significant changes in CO2 properties and impurities in composition, single-sensor technologies such as those based on Coriolis, ultrasonic, differential pressure (DP), acoustic emission (AE) or any other principle alone are not sufficient to provide accurate and robust monitoring across the range of CCS conditions particularly under two-phase gas-liquid flow and transient flow conditions. This research project aims to implement and assess the performance and operability of the multi-modal sensing and machine learning techniques for the monitoring of CO2 under CCS conditions on a dedicated CO2 test facility.
Experimental work was conducted on a CO2 two-phase flow test rig at NCEPU. The first part of the experimental work was to measure the mass flow rate of two-phase CO2. Several sensors and flow instruments, including a Coriolis mass flowmeter (CMF, model Optimass 6400 supplied by Krohne Ltd), a DP transducer that was installed across the Coriolis flowmeter and a capacitive sensor were installed on the test section of the rig, as shown in Figure 1. Various two-phase flow conditions were created under both steady-state and transient flow conditions. For experiments under steady-state flow conditions, the mass flowrate of liquid CO2 is set from 300 kg/h to 3050 kg/h, resulting in the reference gas volume fraction (GVF) from 0% to 87%. Experiments under variable load, start-up and shutdown conditions were conducted to investigate the real-time performance of the instruments and established models. The mass flowrate of gas (liquid) phase CO2 was fixed at 70 kg/h (1500 kg/h) while that of liquid (gas) phase CO2 increased and decreased, resulting in variations in GVF.
The second part of the experimental programme was the detection and monitoring of the leakage from CO2 transportation pipelines under CCS conditions. The experimental setup is shown in Figure 2. Pure CO2 and CO2-N2 mixture in the pipeline were set to leak into the atmosphere by passing the medium through a vent valve. Several tests including different initial conditions were conducted to investigate the release behaviours of CO2 and CO2-N2 mixtures under CCS conditions. AE sensors and temperature transducers are used to acquire signals for the leakage detection and leak rate measurement. Several researchers from NCEPU supported the experimental programme.
Performance of Data Driven Models for Mass Flow Metering of Two-Phase CO2
Mass flow rate, density, capacitance, pressure and temperature data of two-phase CO2 flow were recorded in real time during the tests. Several data driven models based on recurrent neural networks (RNNs) and least squares support vector machines (LS-SVM) were developed to measure the mass flowrate using the data recorded. The relative errors in the predicated mass flow rate are found mostly within ±1.5% under steady-state flow conditions. Figure 3 shows the performance of the data driven model based on RNN under transient flow conditions. The RNN model is capable of tracking the change in CO2 mass flowrate under dynamic CCS conditions with a relative error within ±4%.
Figure 3. Comparison of mass flowrates of CO2 and their relative errors under transient flow conditions
In addition to the data analysis, a study of variable selection for the development of data driven models to improve the measurement accuracy has been undertaken. A gradient boosting random forest (GBRF) model is applied to variable selection for gas-liquid CO2 two-phase flow measurement. The relative errors are within ±1% with the strategically selected variables for the data driven models.
Performance of Data Driven Models for the Leakage Detection of CO2
LS-SVM models coupled with AE sensors and temperature transducers have been developed to measure the leakage rate of CO2 flow from a pipeline and its volume fraction of N2 in the leakage. In a practical leakage, highly pressurized CO2 is released rapidly through the leak hole, which generates AE signals. In order to quantify the leakage, several features are extracted from the raw AE signals such as peak value, peak frequency and AE energy. The temperature transducers measure the temperature drop signals at the leak hole, which is highly correlated with the volume fraction of impurity. The features of the AE signals and the temperature drop are inputs to the LS-LVM model.
The complex flow conditions and significant changes in phase of CO2 pose challenges in the measurement and monitoring of CO2 flow in CCS chains. In this project, we have used a combination of a Coriolis mass flowmeter, a DP transducer and a capacitive sensor with data driven models to measure the CO2 flow rate under both steady-state and transient flow conditions. Several AE sensors are utilized to detect and measure the leakage of CO2 flow a transportation pipeline. The experimental data is still being processed at the time of writing. Variable selection results have recently been published in a leading journal (see a separate blog on UKCCSRC web). The results for CO2 leakage detection and monitoring will be published in the near future.”