Abstract:

In order to accurately analyze the problem of soil temperature characteristics， this paper proposes a prediction model based on attention mechanism of multi-channel long and short-term memory network fused with ARIMA algorithm. The attention-based multi-channel long- and short-term memory network is used to extract important temporal features at different moments of the long and short term， and the ARIMA time series model is used to extract linear features to take advantage of predicting soil temperature more accurately. To validate the proposed model， four evaluation metrics， root mean square error， mean absolute error， mean square error and coefficient of determination， were tested at two weather stations in Switzerland （Laegern and Fluehli weather stations） for the next 6， 12 and 24 hours， while soil depths on 5， 10 and 15 cm soil temperatures. Compared with thse autoregressive integrated moving average model， the long and short-term memory network and the fully connected network， the proposed model has the optimal performance， especially the most significant improvement in the soil temperature model for Fluehli station （10 cm soil depth） during the next 6 hours； the highest relative coefficient of determination value of 0.9965， the lowest root mean square error of 0.3414， the average absolute error of 0.2310 and mean squared error of 0.1165.Therefore， the proposed model can be used as an alternative alternative method for soil temperature estimation.

Key words: machine learning, neural network, soil temperature modeling, attention model, long short-term memory