Scholarlink

<p>We optimized energy use of 185 decision making units (DMUs) under rice production using data envelopment analysis (DEA) followed by development and evaluation of artificial intelligence models for yield prediction. Study evaluated artificial neural networks (ANN) and machine learning (ML) regression models e.g., linear (LR), ridge (RR), Least absolute shrinkage and selection operator (LASSO; L<sub>A</sub>R), elastic net (E<sub>N</sub>R), decision tree (DT), random forest (RF), gradient boosting (GB) and the support vector regression (SVR). Rice production yielded total energy input (E<sub>I</sub>) of 51813.1&#xa0;MJ ha<sup>− 1</sup>; with highest contribution of electricity+ irrigation energy (~ 59.1% of E<sub>I</sub>) and fertilizers (~ 24.5%). DEA elucidated 137 DMUs with technical efficiency scores &lt; 1.00, indicating that ~ 74.1% DMUs were inefficient. It underpins significant (~ 13.1%) energy saving potential; emphasizing excessive use of key inputs particularly irrigation, electricity and fertilizers without proportional yield gains. The best ANN model (10-15-1-1 architecture) yielded high significance (Pearson’s, Kendall’s tau_b and Spearman’s correlation coefficient; R<sub>P</sub> = 0.999**, R<sub>K</sub> = 0.999**, R<sub>S</sub> = 0.999**, R<sup>2</sup> = 0.994**; <i>p</i> &lt; 0.01, respectively), whilst the lowest error (Mean absolute error (MAE) = -4.35E-03, root mean square error (RMSE) = 6.70E-02&#xa0;MJ ha<sup>− 1</sup>, Normalized mean square error (NMSE) = 1.05E-01, mean absolute percent error (MAPE) = -4.05E-01). An analogy for distribution of data-set between measured vs. predicted rice grain yield values (R<sub>P</sub>=0.9999**; training, 0.9196**; validation, 0.9999*; testing and 0.9991** for all data-set; <i>p</i> &lt; 0.01) was observed. Of the ML models, DT outperformed with highest value of R<sup>2</sup> (0.999**; <i>p</i> &lt; 0.01), followed by random forest (R<sup>2</sup> = 0.959**; <i>p</i> &lt; 0.01). The present study identified electricity-driven irrigation and chemical fertilizers inputs as the most significant factors of rice grain yield, together accounting for over 80% of E<sub>I</sub>, while utilizing a foremost control on productivity across DMUs. These findings underpin importance of DEA-based benchmarking for enhancing energy efficiency, while ANNs and DT for précised crop yield prediction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DEA and artificial neural networks for energy optimization and yield prediction of transplanted rice across north-western India

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Gagandeep Kaur,
  • Rajni,
  • Jagtar Singh Sivia

摘要

We optimized energy use of 185 decision making units (DMUs) under rice production using data envelopment analysis (DEA) followed by development and evaluation of artificial intelligence models for yield prediction. Study evaluated artificial neural networks (ANN) and machine learning (ML) regression models e.g., linear (LR), ridge (RR), Least absolute shrinkage and selection operator (LASSO; LAR), elastic net (ENR), decision tree (DT), random forest (RF), gradient boosting (GB) and the support vector regression (SVR). Rice production yielded total energy input (EI) of 51813.1 MJ ha− 1; with highest contribution of electricity+ irrigation energy (~ 59.1% of EI) and fertilizers (~ 24.5%). DEA elucidated 137 DMUs with technical efficiency scores < 1.00, indicating that ~ 74.1% DMUs were inefficient. It underpins significant (~ 13.1%) energy saving potential; emphasizing excessive use of key inputs particularly irrigation, electricity and fertilizers without proportional yield gains. The best ANN model (10-15-1-1 architecture) yielded high significance (Pearson’s, Kendall’s tau_b and Spearman’s correlation coefficient; RP = 0.999**, RK = 0.999**, RS = 0.999**, R2 = 0.994**; p < 0.01, respectively), whilst the lowest error (Mean absolute error (MAE) = -4.35E-03, root mean square error (RMSE) = 6.70E-02 MJ ha− 1, Normalized mean square error (NMSE) = 1.05E-01, mean absolute percent error (MAPE) = -4.05E-01). An analogy for distribution of data-set between measured vs. predicted rice grain yield values (RP=0.9999**; training, 0.9196**; validation, 0.9999*; testing and 0.9991** for all data-set; p < 0.01) was observed. Of the ML models, DT outperformed with highest value of R2 (0.999**; p < 0.01), followed by random forest (R2 = 0.959**; p < 0.01). The present study identified electricity-driven irrigation and chemical fertilizers inputs as the most significant factors of rice grain yield, together accounting for over 80% of EI, while utilizing a foremost control on productivity across DMUs. These findings underpin importance of DEA-based benchmarking for enhancing energy efficiency, while ANNs and DT for précised crop yield prediction.