Multi-time-step ahead daily and hourly intermittent reservoir inflow prediction by artificial intelligent techniques using lumped and distributed data
Multi-time-step ahead daily and hourly intermittent reservoir inflow prediction by artificial intelligent techniques using lumped and distributed data
复制标题
DOI：
10.1016/j.jhydrol.2012.04.045
复制
发表时间：
2012-07
影响因子：
6.4
通讯作者：
V. Jothiprakash;R. Magar
中科院分区：
文献类型：
--
作者：
V. Jothiprakash;R. Magar
在这项研究中，人工智能（AI）技术，如人工神经网络（ANN），自适应神经模糊推理系统（ANFIS）和线性遗传规划（LGP）预测每天和每小时多时间步提前间歇性水库入库流量。为了说明人工智能技术的适用性，间歇Koyna河流域在马哈拉施特拉邦，印度被选为案例研究。根据观测到的日和小时降雨量和水库入库流量的各种类型的时间序列，因果和组合模型开发与集中和分布式输入数据。此外，使用各种性能标准评估模型性能。结果表明，LGP模型的预测性能上级ANN和ANFIS模型，特别是在预测日和小时两个时间步长的峰值流量方面。对总体性能的详细比较表明，组合输入模型（降雨量和流入量的组合）在集中和分布式输入数据建模方面表现更好。据观察，集总输入数据模型的表现略好，因为，除了减少数据中的噪声，更好的技术及其训练方法，网络架构的适当选择，所需的输入，以及数据集的训练测试比。分布式数据的性能稍差是由于较大的变化和较少的观察值。
In this study, artificial intelligent (AI) techniques such as artificial neural network (ANN), Adaptive neuro-fuzzy inference system (ANFIS) and Linear genetic programming (LGP) are used to predict daily and hourly multi-time-step ahead intermittent reservoir inflow. To illustrate the applicability of AI techniques, intermittent Koyna river watershed in Maharashtra, India is chosen as a case study. Based on the observed daily and hourly rainfall and reservoir inflow various types of time-series, cause-effect and combined models are developed with lumped and distributed input data. Further, the model performance was evaluated using various performance criteria. From the results, it is found that the performances of LGP models are found to be superior to ANN and ANFIS models especially in predicting the peak inflows for both daily and hourly time-step. A detailed comparison of the overall performance indicated that the combined input model (combination of rainfall and inflow) performed better in both lumped and distributed input data modelling. It was observed that the lumped input data models performed slightly better because; apart from reducing the noise in the data, the better techniques and their training approach, appropriate selection of network architecture, required inputs, and also training–testing ratios of the data set. The slight poor performance of distributed data is due to large variations and lesser number of observed values.