This article presents a proof-of-concept IoT platform for automated CO₂ emissions monitoring and direct carbon quota calculation at the software level. Instead of physical sensors, a Python-based digital twin was developed to generate synthetic ppm data incorporating sinusoidal oscillations, random noise, and linear drift. The raw values undergo two-point calibration (a = 1.02; b = -5) and are converted to tonnes per hour using the density of CO₂. To assess accuracy, an “ideal” noise- and drift-free sinusoid is generated and compared against the calibrated measurements. The primary purpose of this work is therefore not limited to channel emulation but to validate the feasibility of an end-to-end IoT pipeline covering the entire data lifecycle - from data generation and calibration to quota calculation and registry integration. While the present model is limited to CO₂ as a baseline indicator, it establishes a solid foundation for future extensions toward multi-component mixtures and adsorption dynamics. The results confirm the viability of the proposed “generation → processing → quota calculation → registry” architecture and provide a basis for integration with real IoT devices, MQTT/REST networking protocols, and the national emissions trading system.
In the era of pervasive digitalization, the Internet of Things (IoT) has emerged as a novel paradigm for
monitoring carbon emissions. IoT can be defined as a network of interconnected physical devices,
instruments, and other objects equipped with sensors and software, all linked via the Internet to
collect, store, analyze, and exchange data and derived insights [
Contemporary industrial CO₂ monitoring systems primarily rely on Continuous Emissions
Monitoring Systems (CEMS), which cover roughly 70 % of carbon emissions in the power sector [
Ukraine is currently preparing to implement a national emissions-trading system, as mandated by its Association Agreement with the European Union. This initiative imposes stringent requirements on the transparency, timeliness, and reliability of greenhouse-gas reporting. The absence of an integrated “sensor-to-registry” data transfer mechanism creates a potential gap between on-site quota calculations and their official verification in the state registry. In response, this work develops and tests a proof-of-concept IoT platform featuring a Python-based digital twin of the sensor to generate synthetic CO₂ data, apply two-point calibration, and automatically compute quota volumes for submission to an experimental “mock” registry. The proposed “sensor → quota-calculation → registry” architecture demonstrates the technical feasibility of an end-to-end integration model and provides a foundation for future deployment with physical IoT devices, MQTT/REST protocols, and national reporting systems in the context of Ukraine’s emissions-trading system.
The primary purpose of this study is therefore not limited to channel emulation, but to validate a proof-of-concept IoT pipeline covering the entire data lifecycle – from digital-twin based data generation through calibration and quota calculation to registry integration.
For effective emissions monitoring, intelligent management via the Internet of Things has been
investigated across various sectors – primarily energy, manufacturing, and construction [
A. Arsiwala, F. Elghaish, and M. Zoner (2023) explored pathways to carbon neutrality by
proposing an integrated IoT and AI solution – key components of a digital twin – implemented as an
interactive monitoring dashboard [
Y. Jiang and Z. Mao (2025) note that carbon-emissions monitoring is critical for implementing
reduction strategies, yet excessive reliance on detailed energy data and manual calculations renders
the data-collection process low-frequency, time-lagged, and unreliable. They proposed an
ICEEMDAN-Inception-Transformer model capable of providing accurate hourly carbon-emissions
data collection for energy-sector enterprises [
Li Qingqing et al. (2024) argue that achieving carbon neutrality requires an efficient, reliable carbon ecosystem comprising regulatory bodies, emissions-reduction organizations, and independent auditors. They developed the Modelx+MRV+O system based on IoT and blockchain technologies [12]. Blockchain and IoT can ensure data integrity, transparency, and immutability, facilitating the dissemination of carbon credits within the toolkit of emissions-reduction measures [13], [14].
In our work, we have implemented a software pipeline in pure Python that emulates the complete CO₂ emissions monitoring data lifecycle – “generation → processing → storage → analysis” (Figure 1). In the first stage, the sensor emulator module produces a sequence of ppm readings by modeling ambient concentration as a sinusoidal waveform, overlaid with random Gaussian noise and a linear drift from the initial timestamp. Each data point is tagged with its send time and enqueued into a Python internal queue, which acts as the sole communication channel between the generator and the processor.
(generation of raw ppm with noise, drift, and timestamps) Queu e (queue)
SQLite table ‘readings’
CSV “mock registry” quotas.csv
Calculation of RMSE, MAE, R² Plots of ppm versus ideal Latency plot Processing Engine
Calibration Conversion from ppm to tonnes/hour Latency measurement
The sensor emulator generates a series of NUM_SAMPLES (100) observations at a fixed interval (0.5 s). One hundred measurements provide a statistically significant dataset for metric evaluation, and the 0.5 s interval allows the full dataset to be collected in 50 s while maintaining sufficient resolution to capture the waveform and noise.
For each measurement, the following are computed: 1. The base sinusoid modeling the cyclic variation in concentration is given by: base = 400 + 200 ∙( sin (t - t 0) + 1), 600 (1) where t 0 is the start time of the series.
This value corresponds to ideal_ppm, i.e., ideal_ppm=base. The parameters were selected with the following considerations:
400 ppm – the approximate mean background CO₂ concentration in the atmosphere at ground level.
±200 ppm – the amplitude of cyclic fluctuations, yielding a wave from 200 to 600 ppm; this simulates daily concentration changes resulting, for example, from variations in industrial activity or diurnal photosynthetic uptake by vegetation. This wider span was intentionally chosen to test the robustness of calibration and to approximate possible variations observed in localized industrial or environmental settings.
600 s in the denominator of the sine argument sets the oscillation period to about 10 minutes in the “accelerated” timescale, allowing many daily-like cycles to be emulated within a short measurement session.
2. Random noise, modeled as noise ~ N (0,10) ppm . (2)
A standard deviation of 10 ppm represents a typical level of fluctuation observed in consumer or semi-industrial NDIR sensors over a single measurement session. This magnitude of noise introduces sufficient variability without distorting the overall waveform.
3. Linear drift, defined as (3) drift = t - t 0 ∙ 0.1 ppm / day .
86400 0.1 ppm/day – a small, slow drift typical of NDIR modules caused by temperature variations or sensor aging. It is divided by 86400 s (24 h) so that each second contributes only a minute offset. In the accelerated simulation timescale used in our experiment, this drift is proportionally added to each generated data point, ensuring that long-term sensor instability is represented even within short measurement sessions.
As a result, we obtain the final raw ppm value ppmraw = base + noise + drift together with the send timestamp t send and the elapsed time from the start, re sts = t - t 0.
In the collector, each “message” is read from the queue, the receive time t recv is recorded, and the latency is computed as latency = t recv - t send. Calibration is performed using a two-point method: ppmcorr = a ∙ ppmraw + b, where a = 1.02, b = - 5.
These coefficients were chosen to align two anchor points: when the raw sensor reads 400 ppm, the correction brings it close to the true 400 ppm, and when it reads 1000 ppm, it brings it close to 1000 ppm. A linear regression through these two reference points provides a quick adjustment of the sensor’s output to a calibrated instrument. This simple linear correction compensates for the sensor’s systematic bias.
The corrected ppm values are converted to tonnes per hour using the classical formula: mkg / s = ppmcorr ∙ ρCO2,
106
M t / h = mkg / s ∙ 13060000 .
kg The density of CO₂ (1.977 3 ) is the physical value under standard conditions (1 atm, 25 °C). It m is used to convert concentration (ppm) into a mass flow rate (kg/s). We then apply a factor of 3600/1000 to convert kg/s into tonnes/hour. This step simulates the transformation of concentration into a mass emission rate.
Simultaneously, all processing results are written to a local SQLite database (the readings table for raw and calibrated ppm values and mass flow) for persistent storage and to a CSV file (quotas.csv) (4) (5) (6) (7) acting as a “mock registry” of quotas, thereby simulating integration with the national emissions trading system.
In the final step, the analytics module selectively reads the accumulated data, generates an “ideal” noise- and drift-free sinusoidal curve, and compares it with the calibrated data stream by computing the root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²).
RMS E ppm = √ N i=1 1 ∙ ∑N ( ppmcorri - ideal ppmi)2, 1 N MAE _ ppm = ∑|ppmcorri - ideal _ ppmi|
N i=1 N ∑ ( ppmcorri - ideal _ ppmi)2 R2 _ ppm = 1 - N i=1 ∑ (ideal _ ppmi - ideal _ ppm)2 i=1
, ideal _ ppm = 1 ∑N ideal _ ppmi
N i=1 (8) (9) (10)
These metrics enable the assessment of how closely the synthetic data stream matches the “ideal” by using our noise‐ and drift‐free model based on the same underlying sinusoid but without any random components.
At each step, performance (latency) was measured as the difference between the send timestamp and the in-memory processing time. Concurrently, a latency distribution plot for each message was generated. The number of sent versus received messages (packet-loss) was also calculated.
Thus, the methodology encompasses the entire data lifecycle: parameterized data generation, calibration, quota calculation, validation, and pipeline performance evaluation. This provides a solid foundation for subsequent integration with real IoT devices and the national emissions trading system.
For quantitative evaluation of the accuracy and performance of the developed data pipeline, a series of identical test runs were conducted in a controlled environment. The obtained results enable a direct comparison of the system’s behavior with the theoretical model, free from external latency factors.
These controlled experiments made it possible to evaluate both the accuracy of the synthetic sensor emulation and the stability of the processing pipeline. In particular, we examined how closely the calibrated data follow the ideal sinusoidal pattern and quantified the impact of noise and drift on the overall measurement quality. This setup also allowed us to assess the end-to-end performance of the pipeline in terms of latency, packet loss, and statistical error metrics.
In a 50-second session with a 0.5 s interval, 100 messages were generated and processed. To compare the calibrated measurements against the reference noise- and drift-free sinusoid, Figures 2– 3 present the plot of ppmcorr versus ideal_ppm and the calculated versus ideal CO₂ emissions in tonnes per hour.
As shown in Figure 4, the moving average effectively smooths out high-frequency noise fluctuations, bringing the measured curve closer to the underlying trend while preserving the overall rising shape of the concentration. This confirms the appropriateness of applying filtering to enhance the stability and accuracy of the monitoring algorithm.
The values of the key performance indicators resulting from the simulation are presented in Table 1.
successfully
This table demonstrates that both the modeling accuracy (RMSE, MAE, R²) and the pipeline performance (latency and transmission reliability) remain within the bounds of a software-only emulation. The error distribution (ppmcorr − ideal_ppm) exhibits a mean bias of approximately 0.7 ppm, a median of 1.2 ppm, a standard deviation of 11.1 ppm, a minimum of –25.4 ppm, and a maximum of +29.8 ppm. All 100 messages were processed without any loss (packet loss = 0%) and with effectively zero latency (avg/min/max latency ≈ 0.000 s), underscoring the instantaneous inmemory processing of the Python script.
Figure 5 shows the distribution of processing latency for each of the 100 messages.
As shown in Figure 5, all points lie on the zero line (within the resolution of Python’s timers), further confirming truly instantaneous in-memory processing without any real delays.
These results demonstrate an acceptable level of accuracy for the POC pipeline and its readiness for subsequent deployment with real networking protocols and sensors.
The results make it clear that our software-only emulation of the end-to-end pipeline – from sensor to quota calculation – achieves the desired measurement accuracy and instantaneous data processing, while also highlighting several key areas for future work.
Although the sinusoidal model with added Gaussian noise and linear drift effectively mimics the basic behavior of NDIR sensors, actual devices are subjected to a far broader range of environmental and operational disturbances – temperature swings, humidity, particulate contamination, and electromagnetic interference. In real-world deployments, it will therefore be necessary to implement multi-point calibration or adaptive filtering techniques to correct accumulating errors.
The zero latency observed in our in-memory prototype demonstrates that pure software processing introduces no appreciable delay, but integrating a network layer (whether MQTT or REST) will inevitably add transit delays that depend on link quality and broker load. Industrial practice generally tolerates latencies of 1–2 seconds, so the next step should be to construct a testbed with emulated brokers and measure how these metrics evolve when the system scales to several dozen devices.
Integration with Ukraine’s national emissions trading system (ETS) will require not only a robust data channel but also end-to-end message authentication, encryption, and auditability. A simple API key may be insufficient; public-key infrastructure (PKI) or even a distributed ledger (blockchain) could be considered to guarantee data immutability and trust.
Finally, the economic feasibility of such an automated pipeline must be assessed by weighing the deployment costs of sensors and supporting infrastructure against the time savings and error reductions in reporting. For small and medium-sized enterprises, bundling the core “software + digital twin” solution with outsourced calibration and maintenance services may lower the barrier to adoption.
It should be emphasized that the present proof-of-concept model is intentionally limited to CO₂ as a baseline indicator to validate the feasibility of an end-to-end IoT pipeline for automated quota calculation. In real industrial and environmental conditions, emission streams typically contain multi-component mixtures such as CO₂, CH₄, NOₓ (primarily NO and NO₂), and volatile hydrocarbons, and are subject to competitive adsorption, diffusion, and multiphase equilibrium processes. These phenomena significantly influence both the composition and the effective concentration of emissions. Future research should therefore extend the proposed platform by integrating multi-component models and adsorption dynamics, as highlighted in recent studies on CO₂/CH₄ interactions [15], adsorption modeling on activated carbon [16], and advances in CO₂ capture by absorption and adsorption [17]. Incorporating these aspects will considerably broaden the applicability of the IoT-based monitoring pipeline and make it more representative of real-world scenarios.
In summary, these preliminary findings validate the concept and raise several practical questions – how to adapt the algorithm to real sensors, how to secure a reliable transmission channel and comply with regulatory requirements, and how to develop a cost-effective service model under Ukraine’s ETS. This work thus provides a springboard for subsequent field trials, large-scale deployments, and full integration with physical IoT hardware and national registry systems.
The experimental results validate the effectiveness of the proposed “sensor → processing → storage → analysis” pipeline implemented entirely in software. By employing a Python-based digital twin that generates a sinusoidal baseline overlaid with Gaussian noise and a slow linear drift, we achieved a root-mean-square error (RMSE) of approximately 9 ppm and a coefficient of determination (R²) of about 0.29 following two-point calibration. The complete absence of message loss and the near-zero in-memory processing latency demonstrate the extraordinary speed and reliability of the internal data pipeline.
This proof-of-concept platform lays a solid foundation for further practical and experimental work – especially in light of Ukraine’s forthcoming national emissions trading system (ETS), which demands stringent data timeliness and accuracy. The architecture supports field trials with actual NDIR sensors, encompassing temperature- and humidity-dependent errors and multi-point calibration schemes. Future development will extend the pipeline to real-world deployments by integrating MQTT/REST protocols (introducing realistic network latencies and risks), porting computational modules to microcontrollers, and establishing a secure transmission channel with authentication and encryption for direct ETS registry uploads. Careful economic modeling of platform maintenance and calibration services for small and medium-sized enterprises will be essential to ensure accessibility and cost-effectiveness under resource constraints. In sum, the proposed software pipeline represents a crucial stepping stone toward a full-scale IoT platform for automated CO₂ monitoring and direct quota calculation within state registries.
At the same time, we acknowledge that the present proof-of-concept is limited to CO₂ as a singlecomponent indicator. In real industrial conditions, emission streams contain multi-component mixtures and are affected by competitive adsorption and multiphase equilibrium. Addressing these phenomena in future versions of the platform will further increase its applicability and bring the IoTbased pipeline closer to real-world deployment.
[12] Li Qingqing at all. (2024). An efficient tool for real-time global carbon neutrality with credibility of delicacy management: A Modelx+MRV+O system. Applied Energy. Vol. 372, 123763. https://doi.org/10.1016/j.apenergy.2024.123763. [13] Brunnermeier Markus K., Oehmke, Martin. (2013). The maturity of the financial system. Journal of Finance. 68 (2), 483-521. [14] Boumaiza Ameni, Maher Kenza. (2024). Harnessing Blockchain and IoT for Carbon Credit Exchange to Achieve Pollution Reduction Goals. Energies. 17 (19), 4811. https://doi.org/10.3390/en17194811. [15] Asif, M., Wang, L., Naveen, P., Longinos, S. N., Hazlett, R., Ojha, K., & Panigrahi, D. C. (2024).
Influence of competitive adsorption, diffusion, and dispersion of CH4 and CO2 gases during the CO2-ECBM process. Fuel, 358, 130065. https://doi.org/10.1016/j.fuel.2023.130065. [16] Mo, Z., Li, M., Sun, S., Zhu, R., Zhan, D., Li, A., ... & Yu, Q. (2024). Modeling of activated carbon and multi-scale molecular simulation of its water vapor adsorption: A review. Journal of Environmental Chemical Engineering, 12(5), 113732. https://doi.org/10.1016/j.jece.2024.113732. [17] Soo, X. Y. D., Lee, J. J. C., Wu, W. Y., Tao, L., Wang, C., Zhu, Q., & Bu, J. (2024). Advancements in CO2 capture by absorption and adsorption: A comprehensive review. Journal of CO2 Utilization, 81, 102727. https://doi.org/10.1016/j.jcou.2024.102727.