Daily average relative humidity forecasting via two LSTM-attention methods

The daily average relative humidity is significant for both agriculture and industry. Due to high stochastic, intermittent and non-linear characteristics by nature, the accurate forecasting of the daily average relative humidity is a very challenging task. For improving forecasting performance, two LSTM-attention methods with the attention mechanism added after the input and the attention mechanism added before the output are developed in this paper. First, meteorological data during 1 January 1999 to 31 December 2017 from a meteorological station in Shaanxi, China, were analyzed, where the daily rainfall and daily mean relative humidity were transformed by Log operations to reduce fluctuations and to make their distributions close to the normal distribution. Then, two LSTM-attention methods are designed to forecast the daily average relative humidity, where the LSTM is used to extract time-varying non-linear features and to automatically mine internally causal relationships, and attention mechanisms are applied to improve forecasting accuracy. Experimental results suggest that the method of attention mechanism added after the input gains better performance than the method of attention mechanism added before the output and the baseline LSTM method in MSE, RMSE and MAE.