ARIMA and LSTM based Prediction of Diffused Radiation

Abstract

AbstractThis study demonstrates application of an autoregressive integrated moving average(ARIMA) and long short-term memory (LSTM) network based deep learning optimisers formultistep-ahead prediction of daily mean diffused radiation (DMDR). The DMDR data ofpast 30 years of New Delhi, capital of India has been considered for learning of proposedmodels in python-spyder environment. The study considered four different categories of deeplearning optimisers for prediction using python-spyder platform. The LSTM optimisersconsidered for comparison includes adaptive moment estimation (Adam), adaptive momentestimation with weight regularisation (AdamW), stochastic gradient descent (SGD) and rootmean square propagation (RMSProp). The prediction accuracy of the ARIMA and LSTMbased deep learning optimiser was tested for 10-multistep-ahead predictions using variouserror measurers. The findings clearly highlight superior prediction performance of LSTMbased deep learning models compared to ARIMA model. The LSTM based deep learningmodels yielded lower error responses compared to conventional ARIMA model. The leastmean absolute error (MAE) and mean absolute percentage error (MAPE) values wereobtained using SGD optimiser. Moreover, the least training loss value of 0.0034 was obtainedusing SGD optimiser after 50 epochs.Whereas, the least root mean squared error (RMSE)valueswere obtained using AdamW optimiser. The findings clearly demonstrate suitability ofdeep learning models for precise prediction of environmental attributes.