[COMPLETED]

Multi-Variate Time Series Regression with LSTM for Seoul Bike-Renting Dataset

In this project I work on an LSTM implementation for making a regression prediction on the number of bikes to be made available for rental. The problem is a multi-variate time-series regression supervised problem. Keras with Tensorflow backend used.
For a more thorough discussion of the dataset, architecture choice, methods, design and implementation, see "Discussion.pdf" file
There are two .csv files for the dataset: "SeoulBikeData.csv" is the original file provided in the download (see link below), while "SeoulBikeData2.csv" is a modified version with oneextra column at the end with the month of the datapoint.

Layers

Several architectures were tested with 2, 3, 4 and 5 layers each. The architecture with 3 layers produced the best results, with:
val RMSE = 290.6013
test RMSE = 303.2519
val RAE = 209.757
test RAE = 222.0884

Hyperparameters

Keras Tuner and Hyperband module were employed to find the optimal parameters.
It generated an optimal result of 192 neurons for the first layer, 352 neurons for the second layer and 448 neurons for the third layer. All layers have a bias neuron which is initialised at zero.

Dataset Donwload: http://archive.ics.uci.edu/ml/datasets/Seoul+Bike+Sharing+Demand Eploratory Analysis Notebook based on: https://www.kaggle.com/srvmig/seoulbike