CLIMATE DOWNSCALING AND HYDROLOGICAL IMPACT ASSESSMENT BASED ON LONG SHORT-TERM MEMORY NEURAL NETWORKS

Assessing the impacts of climate change on hydrological systems requires accurate downscaled climate projections. In the past two decades, various statistical and machine-learning techniques have been developed and tested for climate downscaling; however, there is no consensus regarding which technique is the most reliable for climate downscaling and hydrological impact assessment. In this study, an advanced machine-learning technique, Long Short-Term Memory (LSTM) neural network, is used to build multi-model ensembles for downscaling climate projections from a wide ranges of global and regional climate models, and its performance is compared with a number of traditional statistical and machine-learning methods, such as ensemble average, linear regression, Multi-layer Perceptron, Time-lagged Feed-forward Neural Network, and Nonlinear Auto-regression Network with exogenous inputs. The downscaling input consists of temperature and precipitation projections provided by regional climate models, such as CanRCM4, CRCM5, RCA4, and HIRHAM5, and the output is observation data collected from meteorological stations. Performance of the developed LSTM ensemble is evaluated for two case studies in Canada and China. The downscaled climate projections are further used to assess the hydrological impacts in the southwestern mountainous area in China, with the assist of a fully distributed hydrological model, MIKE SHE. The results can support future applications of LSTM neural networks and other similar data-driven techniques for climate downscaling and hydrological impact assessment.