Prediction of time series data using AI technique

This paper will be covering AI techniques in the prediction of climate change data over the course of 140 years. We constructed different neural network architectures to evaluate their effectiveness in forecasting time series data. The various neural network being explored are namely Multi-Layered Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) neural network and a hybrid model which is CNN-LSTM hybrid model. Using time-series feature extraction, this univariate time series problem is transformed into supervised learning problems for which neural networks can be trained on. 70% of the data are split sequentially to form the training data set whilst the 30% of the data are used as test dataset. The metrics that is used to score the models are Root Mean Square Error (RMSE). This FYP report will delve into 2 datasets with one dataset having approximately 2000 rows of data and another bigger dataset with approximately 50,000 rows of data. To increase the accuracy of the models, this paper explores the use of LSTM autoencoders to reconstruct the input space and hyperparameter tuning of the network parameters via grid search. Lastly, the constructed networks are benchmarked across traditional statistical models and machine learning models like random forest, etc in the 2 datasets. It was observed that CNNLSTM hybrid neural network and MLP gives the best result for small dataset whilst linear regression gives the best result for large dataset but in comparison, CNN-LSTM and MLP models are close in accuracy too.Bachelor of Engineering (Information Engineering and Media