Guidance on the construction and selection of relatively simple to complex data-driven models for multi-task streamflow forecasting

With the goal of forecasting streamflow time series with sufficient lead time, we evaluate the efficiency and accuracy of data-based models ranging from relatively simple to complex. Based on this, we systematically explain the model construction and selection process according to lead time, type and amount of data, and optimization method. This analysis involved optimizing the inputs and hyperparameters of four unique data-driven models: Autoregressive Integrated Moving Average (ARIMA), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Transformer (TRANS), which were applied to the Soyang watershed, South Korea. The type and amount of model inputs are determined through a fine-tuning process that samples based on a correlation threshold, correlation to predictand, and autocorrelation to historical data and evaluates the simulated objective function. Hyperparameters are simultaneously optimized using three conventional optimization methods: Bayesian optimization (BO), particle swarm optimization (PSO), and gray wolf optimization (GWO). The experimental results provide insight into the role of input predictors, data preparations (e.g., wavelet transform), hyperparameter optimization, and model structures. From this, we can provide guidelines for model selection. Relatively simple models can be used when the dataset is small or there are few input variables, when only the near future is predicted, or when the selection of optimization methods is limited. However, a more complex model should be selected if the type and amount of data are sufficient, various optimization methods can be applied, or it is necessary to secure more lead time. More parameters, more complex model structures, and more training materials make this possible.