hydropython/One-Step-Head-Daily-Rainfall-Prediction-using-Long-…

In this work, we pursue the three LSTM-Networks architectures in the task of predicting daily rainfall volume using time-series data from Metehara city. We begin by setting parameters to their default value and then use searching strategies to optimize them. Additionally, the study aims to demonstrate the relationship between hyperparameters and metrics as well as their significance. An overview of the paper's main highlights is provided below:

We employ 22 years of daily data from January 1, 1993 to September 18, 2015 to train the models integrating daily rainfall, minimum temperature, maximum temperature, relative humidity, sunshine hour, and wind speed from NMAE (National metrological agency in Ethiopia).

To optimize the hyperparameters in a pre - defined set of search spaces, we employ a random and Bayesian search technique. We follow their significance and correlation to the indicators, which enable us to adjust the parameters effectively.

The best performance hyperparameters information is passed back to the original model of (the LSTM, GRU, and BiLSTM) algorithm to evaluate the model with the updated parameters. The LSTM (Hochreiter and Schmidhuber 1997), GRU(Cho et al. 2014) & BiLSTM (Schuster and Paliwal 1997) models were used in the prediction of daily rainfall. The hyperparameter was searched by Bayesian and random (Bergstra and Bengio 2012) search algorithms. To keep track of hyperparameter, system metrics, and model management of the study Weights & Biases (Wandb) (https://wandb.ai/site) central dashboard was used.