Climate-related variables may not improve monthly scale rainfall predictions by artificial neural networks for the metropolitan region of Belo Horizonte, Brazil

Abstract Artificial neural networks (ANNs) may experience problems due to insufficient or uninformative predictors, and these problems are common for complex predictions such as those for rainfall. However, some studies point to the use of climate variables and anomalies as predictors to make the forecast more accurate. This research aimed to predict the monthly rainfall, one month in advance, in four municipalities of the metropolitan region of Belo Horizonte using an ANN trained with different climate variables; additionally, it aimed to indicate the suitability of such variables as inputs to these models. The models were developed using the MATLAB® software Version R2011a using the NNTOOL toolbox. The ANNs were trained by the multilayer perceptron architecture and the feedforward and backpropagation algorithm using two combinations of input data, with two and six variables, and one combination of input data with the three most correlated variables to observed rainfall from 1970 to 1999 to predict the rainfall from 2000 to 2009. The climate variable most correlated with the rainfall of the following month was the average compensated temperature. Even when using the variables most correlated with precipitation as predictors (0.66 ≤ nt index ≤ 1.26), there was no notable improvement in the predictive capacity of the models when compared to those that did not use climate variables as predictors (0.55 ≤ nt index ≤ 0.80).

摘要：人工神经网络（Artificial Neural Networks，ANNs）可能因预测变量不足或信息匮乏而出现建模性能问题，此类问题在降雨这类复杂预测任务中尤为常见。不过已有研究表明，采用气候变量与气候距平作为预测因子可有效提升预报精度。本研究旨在利用经不同气候变量训练的人工神经网络，对贝洛奥里藏特都会区4个市镇的月降雨量开展提前1个月的预报，同时评估此类变量作为模型输入的适用性。本研究采用MATLAB® R2011a版本软件及NNTOOL工具箱构建模型，以多层感知机架构、前馈反向传播算法训练人工神经网络，共设置3组输入数据组合：含2个变量的组合、含6个变量的组合，以及1组包含与1970-1999年实测降雨量相关性最高的3个变量的组合，以2000-2009年的降雨量作为预报目标。研究结果显示，与次月降雨量相关性最高的气候变量为平均补偿气温。即便采用与降水相关性最高的变量作为预测因子（nt指数范围为0.66≤nt指数≤1.26），与未使用气候变量作为预测因子的模型（nt指数范围为0.55≤nt指数≤0.80）相比，模型的预测性能并未出现显著提升。