A Deep Learning Approach to Predict Weather Data Using Cascaded LSTM Network

Weather prediction is a challenging research problem although the revolutionary advancement in deep learning, along with the availability of big data, has significantly alleviated this problem. Moreover, in terms of robustness and computational cost, this problem has currently interested many researchers to develop numerous models. This paper proposes a lightweight yet powerful deep learning architecture for weather forecasting that can outperform some of the existing well-known models. This architecture mainly uses the LSTM layers in a stacked fashion, with a different number of units in each layer. It takes in multiple weather variables as input features for a given time sequence to forecast the same weather parameters in a multi-input multi-output (MIMO) structure. The resulting models are tested to predict the wind speed, relative humidity, dew point and temperature in this study and experimented with different hyper-parameters consisting a number of LSTM layers, a variable learning rate, number of LSTM units. Two models have been built cascading the basic 1hour-ahead model which predicts the weather parameters for the 2 hours and 3 hours ahead. The obtained results show that the cascaded models perform significantly better than the standard LSTM or 1D convolution networks in shorter period prediction.