Simulate the forecast capacity of a complicated water quality model using the long short-term memory approach

Abstract

The utilization of complicated water quality models is the primary approach used to forecast water quality. These models, however, are not easy to employ because of constraints, such as data limitation, extensive computations, and future boundary conditions. Long short-term memory (LSTM) can overcome such constraints; however, its applications in water quality forecast have rarely been explored. In this study, the ability of LSTM to simulate the forecast capacity of a complicated water quality model (i.e., environmental fluid dynamics code, EFDC) is investigated. First, the EFDC is run to produce a long-term (12 years) time series of six water quality variables. These variables are intrinsically associated with equations embedded in the EFDC that represent the dynamics of the simulated system. The LSTM is developed to forecast the concentration of Chlorophyll a 1–31 d ahead of time using six water quality variables. The generated data are thereafter employed to train a number of LSTMs with different model structures (combinations of input variables, numbers of hidden layers, and lag times). The LSTM performances are evaluated by the Nash–Sutcliffe efficiency coefficient, and random forest is applied to identify the key drivers of LSTM performance. The results show that many LSTMs could achieve an acceptable performance level. Chlorophyll a, water temperature, and total phosphorus are identified as the key drivers of LSTM performances, which are consistent with limnological theories. The number of hidden layers and lag time practically have no impact on the LSTM performance. It is thereby confirmed that an LSTM with a simple structure could simulate the forecast capacity of EFDC. The results also reveal that the mechanism-guided LSTM in our study may capture certain mechanism features. The LSTM is thus expected to be a promising approach for water quality forecast.