⚠ Official Notice: www.ijisrt.com is the official website of the International Journal of Innovative Science and Research Technology (IJISRT) Journal for research paper submission and publication. Please beware of fake or duplicate websites using the IJISRT name.



An Asymmetric Edge-Cloud Decoupled Architecture for Decomposition-Based PM2.5 Forecasting on Resource-Constrained Microcontrollers


Authors : Napatsorn Songsangka

Volume/Issue : Volume 11 - 2026, Issue 8 - August

Google Scholar : https://tinyurl.com/4dyhkamt

DOI : https://doi.org/10.38124/ijisrt/26aug1127

Semantic Scholar

ResearchGate

Note : A published paper may take 4-5 working days from the publication date to appear in PlumX Metrics, Semantic Scholar, and ResearchGate.

Abstract : While hybrid CEEMDAN-LSTM models can improve PM2.5 forecasting accuracy, their ensemble-based decomposition and iterative sifting stages make direct execution of the complete pipeline impractical on resourceconstrained microcontrollers such as ESP32-class edge nodes due to high SRAM requirements. To resolve this computational bottleneck and preserve decomposition benefits, we propose an asymmetric edge-cloud decoupled architecture that assigns CEEMDAN decomposition to the server while physically deploying a lightweight INT8-quantized LSTM on the ESP32-S3 microcontroller. In the proposed online workflow, the server decomposes historical PM2.5 observations and transmits the (168,14) decomposed feature tensor to the ESP32-S3 over Wi-Fi for multi-horizon inference. This data transfer requires a payload of only 2.30 KB per cycle. In the present experiments, server-side feature generation and network transmission were emulated offline, whereas the recurrent inference stage was implemented and executed on the ESP32-S3. Evaluated on a three-year Bangkok dataset, CEEMDAN-LSTM achieves an R² of 0.864 at the 1-hour horizon and 0.595 at 6 hours, outperforming the raw-input baselines at both horizons, with a larger margin at 6 hours. On-device ESP32-S3 evaluation yielded an R² of 0.824 and an average device-side inference latency of 995.48 ms. The serialized INT8 model occupies only 68.38 KB, representing a 54.1% memory reduction compared to the offline Float32 reference without significant accuracy degradation. These results demonstrate the physical feasibility of the recurrent inference stage and support the proposed asymmetric task allocation for multi-scale PM2.5 forecasting on resource-constrained microcontrollers.

Keywords : PM2.5 Forecasting, on-Device Inference, Asymmetric Architecture, CEEMDAN, Long Short-Term Memory (LSTM), TinyML, ESP32-S3, Bangkok Air Quality.

References :

  1. Cohen, A.J., Brauer, M., Burnett, R., Anderson, H.R., Frostad, J., Estep, K., Balakrishnan, K., Brunekreef, B., Dandona, L., Dandona, R., Feigin, V., Freedman, G., Hubbell, B., Jobling, A., Kan, H., Knibbs, L., Liu, Y., Martin, R., Morawska, L., Pope, C.A., Shin, H., Straif, K., Shaddick, G., Thomas, M., van Dingenen, R., van Donkelaar, A., Vos, T., Murray, C.J.L., Forouzanfar, M.H.: Estimates and 25-year trends of the global burden of disease attributable to ambient air pollution: an analysis of data from the Global Burden of Diseases Study 2015. The Lancet. 389, 1907–1918 (2017).
  2. Phairuang, W., Suwattiga, P., Hongtieab, S., Inerb, M., Furuuchi, M., Hata, M.: Characteristics and source apportionment of ambient PM2.5 in Bangkok, Thailand. Atmospheric Pollution Research. 12, 101185 (2021).
  3. Ban, W., Shen, L.: PM2.5 Prediction Based on the CEEMDAN Algorithm and a Machine Learning Hybrid Model. Sustainability (Switzerland). 14, (2022).
  4. Zaini, N., Ean, L.W., Ahmed, A.N., Malek, M.A.: A systematic literature review of deep learning neural network for time series air quality forecasting. Environmental Science and Pollution Research. 29, (2022).
  5. Hochreiter, S., Schmidhuber, J.: Long Short-Term Memory. Neural Comput. 9, 1735–1780 (1997).
  6. Qin, D., Yu, J., Zou, G., Yong, R., Zhao, Q., Zhang, B.: A Novel Combined Prediction Scheme Based on CNN and LSTM for Urban PM2.5 Concentration. IEEE Access. 7, 20050–20059 (2019).
  7. Torres, M.E., Colominas, M.A., Schlotthauer, G., Flandrin, P.: A complete ensemble empirical mode decomposition with adaptive noise. In: 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4144–4147 (2011).
  8. Colominas, M.A., Schlotthauer, G., Torres, M.E.: Improved complete ensemble EMD: A suitable tool for biomedical signal processing. Biomed. Signal Process. Control. 14, 19–29 (2014).
  9. Ameri, R., Hsu, C.C., Band, S.S., Zamani, M., Shu, C.M., Khorsandroo, S.: Forecasting PM 2.5 concentration based on integrating of CEEMDAN decomposition method with SVM and LSTM. Ecotoxicol. Environ. Saf. 266, 115572 (2023).
  10. Alajlan, N.N., Ibrahim, D.M.: TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications. Micromachines (Basel). 13, (2022).
  11. Abadade, Y., Temouden, A., Bamoumen, H., Benamar, N., Chtouki, Y., Senhaji Hafid, A.: A Comprehensive Survey on TinyML. IEEE Access. 11, 96892–96922 (2023).
  12. David, R., Duke, J., Jain, A., Reddi, V.J., Jeffries, N., Li, J., Kreeger, N., Nappier, I., Natraj, M., Regev, S., Rhodes, R., Wang, T., Warden, P.: TensorFlow Lite Micro: Embedded machine learning on TinyML systems. In: Proceedings of the 4th MLSys Conference. , San Jose, CA (2021).
  13. Warden, P., Situnayake, D.: TinyML Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. O’Reilly Media, CA (2019).
  14. Cai, P., Zhang, C., Chai, J.: Forecasting hourly PM2.5 concentrations based on decomposition-ensemble-reconstruction framework incorporating deep learning algorithms. Data Science and Management. 6, 46–54 (2023).
  15. Teng, M., Li, S., Xing, J., Song, G., Yang, J., Dong, J., Zeng, X., Qin, Y.: 24-Hour prediction of PM2.5 concentrations by combining empirical mode decomposition and bidirectional long short-term memory neural network. Science of the Total Environment. 821, (2022).
  16. Sun, C., Li, J., Sulaiman, R., Alotaibi, B.S., Elattar, S., Abuhussain, M.: Air Quality Prediction and Multi-Task Offloading based on Deep Learning Methods in Edge Computing. J. Grid Comput. 21, (2023).
  17. Aman, N., Manomaiphiboon, K., Pengchai, P., Assareh, N., Dejchanchaiwong, R.: High-resolution spatial modeling of traffic-related air pollution in Bangkok, Thailand. Atmos. Environ. 246, 118090 (2021).
  18. Dhammapala, R.: Analysis of fine particle pollution data measured at 29 US diplomatic posts worldwide. Atmos. Environ. 213, 367–376 (2019).
  19. Huang, N., Shen, Z., Long, S., Wu, M., Shih, H., Zheng, Q., Yen, N.-C., Chao Tung, C., Liu, H.: The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences. 454, 903–995 (1998).
  20. Kingma, D., Ba, J.: Adam: A Method for Stochastic Optimization. In: International Conference on Learning Representations (2015).
  21. Espressif Systems: ESP-IDF Programming Guide: Partition Tables, https://docs.espressif.com/projects/esp-idf/, last accessed 2026/07/11.

While hybrid CEEMDAN-LSTM models can improve PM2.5 forecasting accuracy, their ensemble-based decomposition and iterative sifting stages make direct execution of the complete pipeline impractical on resourceconstrained microcontrollers such as ESP32-class edge nodes due to high SRAM requirements. To resolve this computational bottleneck and preserve decomposition benefits, we propose an asymmetric edge-cloud decoupled architecture that assigns CEEMDAN decomposition to the server while physically deploying a lightweight INT8-quantized LSTM on the ESP32-S3 microcontroller. In the proposed online workflow, the server decomposes historical PM2.5 observations and transmits the (168,14) decomposed feature tensor to the ESP32-S3 over Wi-Fi for multi-horizon inference. This data transfer requires a payload of only 2.30 KB per cycle. In the present experiments, server-side feature generation and network transmission were emulated offline, whereas the recurrent inference stage was implemented and executed on the ESP32-S3. Evaluated on a three-year Bangkok dataset, CEEMDAN-LSTM achieves an R² of 0.864 at the 1-hour horizon and 0.595 at 6 hours, outperforming the raw-input baselines at both horizons, with a larger margin at 6 hours. On-device ESP32-S3 evaluation yielded an R² of 0.824 and an average device-side inference latency of 995.48 ms. The serialized INT8 model occupies only 68.38 KB, representing a 54.1% memory reduction compared to the offline Float32 reference without significant accuracy degradation. These results demonstrate the physical feasibility of the recurrent inference stage and support the proposed asymmetric task allocation for multi-scale PM2.5 forecasting on resource-constrained microcontrollers.

Keywords : PM2.5 Forecasting, on-Device Inference, Asymmetric Architecture, CEEMDAN, Long Short-Term Memory (LSTM), TinyML, ESP32-S3, Bangkok Air Quality.

Paper Submission Last Date
30 - September - 2026

SUBMIT YOUR PAPER CALL FOR PAPERS
Video Explanation for Published paper

Never miss an update from Papermashup

Get notified about the latest tutorials and downloads.

Subscribe by Email

Get alerts directly into your inbox after each post and stay updated.
Subscribe
OR

Subscribe by RSS

Add our RSS to your feedreader to get regular updates from us.
Subscribe