Utilization of Kernel Function/Decision Tree-Based Soft-Computing Approaches for Precise Forecasting of Methane Production from the Co-Digestion of Cow Manure and Slaughterhouse Waste

Ali M. M., Yetilmezsoy K., Makoui S., Ebeya C. C., Azqhandi M. H. A., Yahya A. M.

8th International African Conference on Contemporary Scientific Research, Nairobi, Kenya, 11 - 13 Aralık 2024, ss.62-78

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Nairobi
  • Basıldığı Ülke: Kenya
  • Sayfa Sayıları: ss.62-78
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

The present computational research aimed to assess the effectiveness of various soft-computing techniques in predicting methane production from the co-digestion of cow manure and liquid slaughterhouse waste using five input variables including digestion time (up to 3793 h), biogas production (896.47 ± 395.77 mL/g VS), pH (6.37 ± 0.28), biogas temperature (39.99 ± 8.58 °C), and digestion temperature (35.07 ± 8.50 °C) for a total of 4548 data records. The results from the experiment regarding the biochemical methane potential (BMP) of this recalcitrant substrate demonstrated that the methane (CH4) production began after 37 hours (1.54 days) of incubation. Efficient conversion of volatile solids (VS) into CH4 was exemplified by the biomethanation process which yielded 673.44 mL CH4/g VS. The CH4 content at steady state was recorded as 56% within the biodegradation conditions spanning from upper mesophilic to lower thermophilic temperatures. Two kernel function-based methods and three decision tree-based approaches were jointly utilized for the first time to forecast CH4 production from a particular substrate combination instead of using traditional and often error-prone sigmoidal microbial growth curve models. Predictive accuracy was evaluated based on more than 30 different statistical indices. Performance evaluation metrics for the testing set confirmed that the random forest (RF)-based model outperformed other methods in forecasting the biomethanation process behavior (R2 = 0.9997, MAE = 2.7384 mL CH4/g VS, RMSE = 3.9317 mL CH4/g VS, WI = 0.9999, NSE = 0.9997, LMI = 0.9859). Based on the statistical analyses, it was observed that the novel RF-based approach obtained the least uncertainty band of (±1.96Se = 7.7229 mL CH4/g VS) and the smallest 95% confidence level estimation uncertainty (U95 = 29.3238 mL CH4/g VS). Out of several models developed, the random forest (RF)-based model achieved the least corrected Akaike information criterion (AICc = 633.9151) and attained the highest overall accuracy (ψ = 6.8978), indicating its effectiveness and accuracy over other competing models.