基于LSTM-Attention-GAN的潜在蒸散量预测研究

Prediction of Potential Evapotranspiration Based on LSTM-Attention-GAN

摘要: 全球月尺度潜在蒸散量(PET)预测面临驱动因素多源耦合、时序依赖复杂等挑战；同时，传统以MSE、MAE等点对点损失为主的回归方法容易出现均值收缩，进而影响预测分布对尾部波动的表达。基于1981~2020年、1˚ × 1˚统一网格数据，文章构建了“过去24个月多源气象–水文变量预测下一月PET”的序列学习框架。模型以长短期记忆网络(LSTM)为主干，引入时间注意力与特征注意力以强化关键月份与关键驱动因子的表征；进一步结合生成对抗网络(GAN)，在分布层面对预测结果施加约束，以促进预测分布与观测分布的一致性。实验结果表明，引入注意力机制后RMSE显著降低且系统偏差减小；在此基础上加入对抗学习后，模型在分布形态相似性、综合一致性及高值区偏差控制方面进一步改善，整体表现更为稳健。

Abstract: The prediction of global monthly potential evapotranspiration (PET) faces challenges such as multi-source coupling of driving factors and complex time series dependence. At the same time, the traditional regression methods based on point-to-point losses such as MSE and MAE are prone to mean shrinkage, which in turn affects the expression of tail fluctuations in the prediction distribution. Based on the 1˚ × 1˚ unified grid data from 1981 to 2020, this paper constructs a sequence learning framework of “multi-source meteorological-hydrological variables in the past 24 months predict PET in the next month”. The model takes the long-term and short-term memory network (LSTM) as the backbone, and introduces time attention and feature attention to strengthen the representation of key months and key driving factors. Furthermore, the generative adversarial network (GAN) is combined to impose constraints on the prediction results at the distribution level to promote the consistency between the prediction distribution and the observation distribution. The experimental results show that the RMSE is significantly reduced, and the system deviation is reduced after the attention mechanism is introduced. On this basis, after adding adversarial learning, the model is further improved in terms of distribution similarity, comprehensive consistency, and deviation control in high-value areas, and the overall performance is more robust.