Application of Regression and Artificial Neural Network in Ground Temperature Processing

Different strategies have been researched using various artificial intelligence techniques in reported meteorological data processing systems in recent decades. Nowadays, various artificial intelligence algorithms had been employed to process meteorological data around the world. Accordingly, in this study, using the methods including linear regression (LR) and artificial neural network (ANNs) to analyze daily and yearly meteorological parameters to predict ground temperature in Chengdu. The data used in LR and ANN came from Wen Jiang reference meteorological station in Chengdu, China from 1968 to 2017 (50 years). A three-layer feed-forward network (FFNN) was built for the data processing and a backpropagation algorithm (BP) was also included. In order to get an effective and accurate prediction result, the correlation coefficients between the meteorological parameters (mean ground temperature, mean atmospheric temperature, mean relative humidity, mean wind speed, mean rainfall) were calculated taking them two by two. The input parameters which were processed data that included mean ground temperature and mean atmospheric temperature employed to train the neural networks. The output layer composed of one neuron to represent the prediction of ground temperature. Moreover, correlation coefficient (R), error between predicted value and actual value (error), Root Mean Squared Error (RMSE) and coefficient of determination (R2) are the standard sticks adopted for model developing and data predicting accuracy measurement. The data for the first 40 years was used as the training group, and the data for the last 10 years was used as the test group. The results show that the ANNs method can predict the ground temperature in a small error range and it performs better than the LR method. Additionally, results could be better, using relative less neurons that reduces training time and complexity.