Runoff prediction of urban stream based on the discharge of pump stations using improved multi-layer perceptron applying new optimizers combined with a harmony search

Prediction using neural networks is applied in various fields, including hydrology and water resource manage-ment. In the learning process, an optimizer is used to determine the relationship (weights and biases) between the input and output data to lower the value of the loss function. Existing optimizers rely absolutely on the relationship that are initially generated and have the probability to converge to a local optimum. Additionally, since there is no structure for storing information about the previously generated correlation, there is a proba-bility that the optimal weights and biases could not be found even when learning proceeds. It is necessary to apply a meta-heuristic optimization algorithm that can consider both global/local search and has a structure that can store the previously generated solution. In this study, a new optimizer was combined with a harmony search (HS), to improve existing optimizers. To test the performance of the improved multi-layer perceptron using the new optimizer (IMLP), the runoff of the Dorim stream in Seoul was predicted. The rainfall and the discharge of pump stations were constructed as input data, and data preprocessing was performed by normalization and correlation coefficients. The IMLP showed improved accuracy and could be used to manage urban streams.