ANFIS against LSTM: precipitation time series forecasting

Precipitation forecasting has risen to importance in recent years due to its complexity and enduring uses, such as flood forecasts. Existing models rely on complicated statistical models that are often computationally and financially inefficient or could not be used in downstream applications. As a result, techniques that combine artificial intelligence with time-series data are being investigated to address these limitations. This study aimed to compare adaptive neuro-fuzzy inference system (ANFIS) and long short term memory (LSTM) networks as a representation of machine learning and deep learning, respectively. A typical case study of the lower Mississippi River Basin monthly time series precipitation has been successfully predicted using the ANFIS and LSTM network. The ANFIS and the LSTM networks have been run as the same as possible setup. Good practices have been implemented concerning better prediction, such as data standardization and a hybrid method of solver and optimization. LSTM indicates betterresults concerning ۱.۸۶۵, ۱.۳۶۶, and ۰.۳۶۶ for MSE, RMSE, and MAPE, respectively. Results demonstrate ۴۶.۹۸%, ۲۷.۱۸%, ۲۷.۶۹%, and ۵۱.۳۳% improvement in MSE,RMSE, MAPE, and Time of training for the LSTM network concerning ANFIS.

Keywords:

LSTM , ANFIS , time series , precipitation , deep learning , machine learning