Vocational School of Technical Sciences, Department of Environmental Protection Technologies, Akdeniz University, Antalya, Turkey.
Department of Civil Engineering, University of Sistan and Baluchestan, Zahedan, Iran.
Environ Sci Pollut Res Int. 2024 Jun;31(30):42719-42749. doi: 10.1007/s11356-024-33987-3. Epub 2024 Jun 15.
Accurately predicting potential evapotranspiration (PET) is crucial in water resource management, agricultural planning, and climate change studies. This research aims to investigate the performance of two machine learning methods, the adaptive network-based fuzzy inference system (ANFIS) and the deep belief network (DBN), in forecasting PET, as well as to explore the potential of hybridizing the ANFIS approach with the Snake Optimizer (ANFIS-SO) algorithm. The study utilized a comprehensive dataset spanning the period from 1983 to 2020. The ANFIS, ANFIS-SO, and DBN models were developed, and their performances were evaluated using statistical metrics, including R, , NSE, WI, STD, and RMSE. The results showcase the exceptional performance of the DBN model, which achieved R and values of 0.99 and NSE and WI scores of 0.99 across the nine stations analyzed. In contrast, the standard ANFIS method exhibited relatively weaker performance, with R and values ranging from 0.52 to 0.88. However, the ANFIS-SO approach demonstrated a substantial improvement, with R and values ranging from 0.94 to 0.99, suggesting the value of optimization techniques in enhancing the model's capabilities. The Taylor diagram and violin plots with box plots further corroborated the comparative analysis, highlighting the DBN model's superior ability to closely match the observed standard deviation and correlation and its consistent and low-variance predictions. The ANFIS-SO method also exhibited enhanced performance in these visual representations, with an RMSE of 0.86, compared to 0.95 for the standard ANFIS. The insights gained from this study can inform the selection of the most appropriate modeling technique, whether it be the high-precision DBN, the flexible ANFIS, or the optimized ANFIS-SO approach, based on the specific requirements and constraints of the application.
准确预测潜在蒸散量（PET）在水资源管理、农业规划和气候变化研究中至关重要。本研究旨在探讨两种机器学习方法，自适应网络模糊推理系统（ANFIS）和深度置信网络（DBN）在预测 PET 方面的性能，以及探索将 ANFIS 方法与 Snake Optimizer（ANFIS-SO）算法相结合的潜力。该研究使用了一个涵盖 1983 年至 2020 年期间的综合数据集。开发了 ANFIS、ANFIS-SO 和 DBN 模型，并使用统计指标，包括 R²、、NSE、WI、STD 和 RMSE，对它们的性能进行了评估。结果展示了 DBN 模型的卓越性能，该模型在分析的九个站点中，R²和 的值分别达到 0.99 和 NSE 和 WI 分数为 0.99。相比之下，标准的 ANFIS 方法表现相对较弱，R²和 的值范围为 0.52 到 0.88。然而，ANFIS-SO 方法表现出了显著的改进，R²和 的值范围为 0.94 到 0.99，表明优化技术在增强模型能力方面的价值。泰勒图和小提琴图与箱线图进一步证实了比较分析，突出了 DBN 模型更接近观测标准偏差和相关性的能力，以及其一致和低方差的预测。ANFIS-SO 方法在这些可视化表示中也表现出了改进的性能，其 RMSE 为 0.86，而标准的 ANFIS 为 0.95。从这项研究中获得的见解可以根据应用的具体要求和限制，为选择最合适的建模技术提供信息，无论是高精度的 DBN、灵活的 ANFIS 还是优化的 ANFIS-SO 方法。
