Automated lag-selection for multi-step univariate time series forecast using Bayesian Optimization: Forecast station-wise monthly rainfall of nine divisional cities of Bangladesh

Abstract: Rainfall is an essential hydrological component, and most of the economic activities of an agrarian country like Bangladesh depend on rainfall. An accurate rainfall forecast can help make necessary decisions and reduce the damages caused by heavy or low to no rainfall. The monthly average rainfall is a time series data, and recently, long short-term memory (LSTM) neural networks are being used heavily for time series forecasting problems. One major challenge of forecasting using LSTMs is to select the appropriate number of lag values. In this research, we considered the number of lag values selected as a hyperparameter of LSTM; it, with the other hyperparameters determining LSTMs structure, has been optimized using Bayesian optimization. We used our proposed method to forecast rainfall for nine different weather stations of Bangladesh. Finally, the performance of the proposed model has been compared with some other LSTM with different lag-selection methods and some several popular machine learning and statistical forecasting models.