三种水质动态预测模型在米山水库的应用与结果对比

Comparison of Three Models for PredictingWater Quality in Mishan Reservoir

【目的】分析不同水质预测模型的预测精度,探寻最优的水库水质预测方法。【方法】分别构建了季节性差分自回归滑动平均模型(SARIMA)、霍尔特-温特(Holt-Winters)模型和长短时记忆(LSTM)神经网络模型,利用米山水库2012—2018年的月平均电导率观测数据对模型进行训练,利用2019年月电导率实测数据对模型进行验证,考察3种预测模型的准确性和稳定性。【结果】SARIMA模型和Holt-Winters模型仅能考察水质数据的时序演化趋势,预测精度较低;相比之下,LSTM神经网络模型能同时考察水质数据的时序演化趋势及不同时刻之间的前后依赖关系,具有较强的非线性映射能力,预测精度最高。【结论】LSTM神经网络预测模型仅在电导率值突变处误差相对较大,但整体预测效果较为理想,因此在水质预测中更加具有推广价值。

【Objective】Accurate prediction of water quality is of great importance for protecting water ecological environment and improving water resource management,but difficult due to the combined influence of various factors including human activities,rainfall,temperature and hydrodynamic conditions which are complex and uncertain.In this paper we compare three models,the SARIMA model,Holt-Winters model and LSTM neural network model,for predicting water quality of reservoir.【Method】Electric conductivity of water was used as a proxy for water quality.Data measured monthly from 2012—2018 from Mishan Reservoir was used for model training,and the data measured in 2019 was used to test the models.【Result】The SARIMA and Holt-Winters models are comparatively simple,but because they only consider time series of water quality data,their accuracy is low.In contrast,the LSTM neural network model considers factors that affect water quality and implicitly represents the nonlinearity of the factors in their impact on water quality,it is more accurate than other two models.【Conclusion】In general,the LSTM neural network model is reliably,giving rise to large error only when there were sudden changes in the conductivity.Overall,it is reliable and accurate for predicting change in water quality induced by variations in the environment.