Urban land surface temperature prediction using parallel STL-Bi-LSTM neural network

Accurate temperature prediction is of great significance to human life and social economy. A series traditional methods machine learning have been proposed achieve prediction, but it still a challenging problem. We propose model that combines seasonal trend decomposition using loess (STL) the bidirectional long short-term memory (Bi-LSTM) network high-accuracy daily average China cities. The decomposes data STL into component, remainder component. Decomposition components original are input two-layer Bi-LSTM learn features data, sum three result added learnable weights as result. experimental results show root mean square error absolute on testing 0.11 0.09, respectively, which lower than 0.35 0.27 STL-LSTM, 2.73 2.07 EMD-LSTM, 0.39 0.15 STL-SVM, achieving higher precision prediction.