Metaheuristic Techniques Optimised LSTM Network for Improved Weather Prediction

Tremendous amount of meteorological data is being generated on a daily basis from a number of sources such as weather stations, balloons, satellites, sensors etc. Timely weather prediction helps people plan everyday life events. Long Short Term Memory (LSTM) networks perform extremely well for capturing dependencies in time series datasets. Number of neurons in hidden layer highly impacts the performance of the LSTM network. The hit and Trial method for selecting the neurons consumes a lot of time and resources and might not lead to a global optimum solution. In this research work, two techniques, i. e. Genetic Algorithm optimized LSTM (LSTMGA) and Artificial Bee Colony optimized LSTM (LSTMABC) have been proposed and implemented to automate the selection of the hidden layer's neurons for improved weather prediction. Proposed techniques have been implemented using DeepLearning4j library making the techniques scalable on clusters of machines. For evaluating the performance of proposed techniques, 15 years Brazil's weather dataset has been used. The experimental results proved that the proposed LSTMGA and LSTMABC techniques have the reduced MAE value of 0. 0074 and 0. 0075 respectively while the LSTM network without any optimization has a relatively higher MAE value of 0. 0086.