基于多种学习方案LSTM的信江流域径流预测

Runoff Prediction of Xinjiang River Basin Based on Multiple Learning Schemes LSTM

长短期记忆神经网络(LSTM)在径流预测中具有广泛应用,不同的输入使神经网络具有不同的学习方案,从而影响到模型性能。设置3种不同的LSTM学习方案,以前期径流预测当日径流(方案一)、以前期降雨预测当日径流(方案二)和以前期径流和前期降雨预测当日径流(方案三),比较其在相同模型结构下对信江流域丰水期和枯水期径流预测的性能。结果表明,丰水期和枯水期时方案三拟合度最高,平均绝对误差为0.012 6和0.007 6,纳什效率系数为0.94和0.96,对于信江流域基于LSTM的日径流预测,应当将前期降雨与前期径流结合起来作为模型输入。研究对基于数据驱动的径流预测输入集数据的选取有参考价值。

Long short-term memory(LSTM) has been widely used in runoff prediction. Different inputs make the LSTM have different learning schemes, which affects the performance of the model. Three different LSTM learning schemes are set up, including Scheme 1-taking previous runoff as inputs, Scheme 2-taking previous rainfall as inputs, and Scheme 3-taking both previous runoff and previous rainfall as inputs. The performances of three schemes in predicting runoff during wet and dry periods in Xinjiang River Basin under the same model structure are compared. The results show that, in both the wet and dry periods, Scheme 3 have highest degree of fitting, and the final mean absolute error(MAE) are 0.012 6 and 0.007 6 respectively and the Nash-Sutcliffe efficiency coefficient(NSE) are 0.94 and 0.96 respectively. For the daily runoff prediction of Xinjiang River Basin based on LSTM, the previous rainfall and previous runoff should be combined as inputs. The research has guiding and reference value for the selection of input data of data-driven watershed runoff prediction.