Analysis of Statistical and Deep Learning Techniques for Temperature Forecasting

Abstract

In the field of meteorology, temperature forecasting is a significant task as it has been a key factor in industrial, agricultural, renewable energy, and other sectors. High accuracy in temperature forecasting is needed for decision-making in advance. Since temperature varies over time and has been studied to have non-trivial long-range correlation, non-linear behavior, and seasonal variability, it is important to implement an appropriate methodology to forecast accurately. In this paper, we have reviewed the performance of statistical approaches such as AR and ARIMA with RNN, LSTM, GRU, and LSTM-RNN Deep Learning models. The models were tested for short-term temperature forecasting for a period of 48 hours. Among the statistical models, the AR model showed notable performance with a r2 score of 0.955 for triennial 1 and for the same, the Deep Learning models also performed nearly equal to that of the statistical models and thus hybrid LSTM-RNN model was tested. The hybrid model obtained the highest r2 score of 0.960. The difference in RMSE, MAE and r2 scores are not significantly different for both Statistical and Vanilla Deep Learning approaches. However, the hybrid model provided a better r2 score, and LIME explanations have been generated for the same in order to understand the dependencies over a point forecast. Based on the reviewed results, it can be concluded that for short-term forecasting, both Statistical and Deep Learning models perform nearly equally.