China’s freshwater resources are relatively small per capita, and the traditional passive control of hydrographic outbreaks can no longer meet modern water management needs. Data-driven models, such as Long Short-Term Memory Networks (LSTMs), have been gradually applied to water resources management, but most of the research has focused on the enhancement of the prediction effect of hybrid models while neglecting the importance of data structure. In this study, we predicted the number of dominant algae (blue-green algae) in a water source based on LSTM and explored the effects of different feature combinations and time window steps on the prediction performance. It was found that the model prediction was significantly improved by adding multiple features, and the R2 improved by 31.98% compared with single feature prediction. Meanwhile, as the time window (T-value) increased from 7 to 300, the R2 improved by 0.4%, but the iteration time increased by 96%. The results suggested that appropriate input feature selection is beneficial for model prediction, while longer time windows led to reduced model prediction benefits. Lastly, this study offers insights into future research directions from three key dimensions: the input indicator, optimization algorithm, and model combination.
Read full abstract