Developing a novel hybrid model based on deep neural networks and discrete wavelet transform algorithm for prediction of daily air temperature

Abstract

The precise predicting of air temperature has a significant influence in many sectors such as agriculture, industry, modeling environmental processes. In this work, to predict the mean daily time series air temperature in Muğla city (AT_m), Turkey, initially, five different layer structures of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning-based neural network models through the seq2seq regression forecast module are developed. Then, based on performance evaluation metrics, an optimal DL-based layer network structure designed is chosen to hybridize with the wavelet transform (WT) algorithm (i.e., WT-DNN model) to enhance the estimation capability. In this direction, among potential meteorological variables considered, the average daily sunshine duration (SSD) (hours), total global solar radiation (TGSR) (kw. hour/m²), and total global insolation intensity (TGSI) (watt/m²) from Jan 2014 to Dec 2019 are picked as the most effective input variables through correlation analysis to predict AT_m. To thwart overfitting and underfitting problems, different algorithm tuning along with trial-and-error procedures through diverse types of hyper-parameters are performed. Consistent with the performance evaluation standards, comparison plots, and Total Learnable Parameters (TLP) value, the state-of-the-art and unique proposed hybrid WT-(LSTM × GRU) model (i.e., hybrid WT with the coupled version of LSTM and GRU models via Multiplication layer (\(\times\))) is confirmed as the best model developed. This hybrid model under the ideal hyper-parameters resulted in an R² = 0.94, an RMSE = 0.56 (^°C), an MBE = -0.5 (^°C), AICc = -382.01, and a running time of 376 (s) in 2000 iterations. Nonetheless, the standard single LSTM layer network model as benchmark model resulted in an R² = 0.63, an RMSE = 4.69 (^°C), an MBE = -0.89 (^°C), AICc = 1021.8, and a running time of 186 (s) in 2000 iterations.