Use of LSTM Machine Learning Technique to Forecast Rainfall

Abstract

The current study attempted to predict rainfall on a monthly basis using sequential Long short-term memory (LSTM) algorithm for four selected stations (Kohima, Mon, Phek and Wokha) of Nagaland, India. The LSTM model was built using Python, with data split 80% for training and 20% for testing. Numerous trials with varying LSTM hyperparameters were conducted to identify the best-performing model.

For monthly rainfall predictions, the model showed strong performance, the value of MAE, MSE, RMSE, NSE ranged from 0.228 to 0.241, 0.01 to 0.014, 0.103 to 0.117, and 0.755 to 0.796 respectively during training and 0.196 to 0.245, 0.009 to 0.013, 0.097 to 0.116, and 0.71 to 0.75 respectively during testing. The near-zero errors values (MAE, MSE, RMSE) and NSE values close to 1.0 indicated the model’s suitability for making reliable predictions.

Ultimately, using the developed sequential LSTM model, monthly rainfall was forecasted for the selected four stations from the year 2023 to 2026.