1.School of Automation,Nanjing University of Information Science and Technology;2.Hubei public meteorological service center
日降水量变化呈现显著的非线性特征,对其进行精准预测难度很大。近年来,长短时记忆网络(Long-Short Term Memory, LSTM)在降水量预测中优势明显。然而,LSTM的深层结构造成了其存在过拟合、时滞等缺点,从而影响预测精度。注意到宽度学习系统(Broad Learning System, BLS)是直接计算权重,其无需多次迭代的特点可以帮助解决LSTM存在的缺点。但噪声和异常值对建模依旧有着不良影响。基于此,提出改进的加权宽度学习系统(Weighted Broad Learning System, WBLS)。利用加权惩罚因子约束每个样本,分别为正常和异常样本分配高和低权重来增加和减少它们的影响。本文结合了LSTM和WBLS的优点提出LSTM-WBLS日降水量预测模型。选取湖北省巴东站近20年日降水量实测数据进行实证研究,且考虑气压、气温、湿度、风速和日照等因素。结果表明,与现有的预测模型相比,LSTM-BLS模型在所有评价指标上均预测精度最高。特别地,WBLS模块加入解决了LSTM的时滞性问题。而且在不同时间步长下,新模型预测精度亦表现最佳,证明了其稳定性。在运算效率上,LSTM-WBLS和LSTM相比,并未降低。
The change of daily precipitation presents significant non-linear characteristics, and it is very difficult to accurately predict it. In recent years, Long-Short Term Memory (LSTM) has obvious advantages in precipitation prediction. However, the deep structure of LSTM causes its shortcomings such as over-fitting and time lag, which affect the prediction accuracy. Note that the Broad Learning System (BLS) directly calculates weights, and its feature of not requiring multiple iterations can help solve the shortcomings of LSTM. However, noise and outliers still have an adverse effect on modeling. Based on this, an improved Weighted Broad Learning System (WBLS) is proposed. Use a weighted penalty factor to constrain each sample, and assign high and low weights to normal and abnormal samples to increase and decrease their influence. This paper combines the advantages of LSTM and WBLS to propose a LSTM-WBLS daily precipitation prediction model. Empirical research is carried out on the actual data of daily precipitation at Badong Station in Hubei Province in the past 20 years, and factors such as air pressure, temperature, humidity, wind speed and sunshine are taken into consideration. The results show that, compared with the existing prediction models, the LSTM-BLS model has the highest prediction accuracy in all evaluation indicators. In particular, the WBLS module has been added to solve the time lag problem of LSTM. Moreover, under different time steps, the prediction accuracy of the new model also performed best, proving its stability. In terms of computational efficiency, LSTM-WBLS has not decreased compared with LSTM.
南京信息工程大学学报 ® 2022 版权所有 技术支持：北京勤云科技发展有限公司