sindhri / airplane_passenger_prediction

Use 144 months of historical data to predict the number of passengers (in thousands) in an international airplane for any given month. Simple MLP and MLP with a larger window are used and compared.

Airline Passenger Prediction

Predict the number of international airline passengers in units of 1,000, given a year and a month. The data ranges from January 1949 to December 1960 or 12 years, with 144 observations, one for each month.
Credit: Deep Learning with Python by Jason Brownlee
data source: https://raw.githubusercontent.com/jbrownlee/Datasets/master/airline-passengers.csv

EDA

Plot the 144 observations and it has a prominent cycle in it.

Data preparation

Split the time series sampple in 67% for training and 33% for validation

Models

1. MLP lookback = 1

1. Create a function, which takes argument X and 1 at time t and output Y at time t+1 based on the dataset
2. Use the function to create training and prediction from the dataset
   Train Score: 23.03 RMSE
   Test Score: 48.49 RMSE
   The model has anaverage error of 23 passengers (in thousands) on the training dataset and 48 passengers (inthousands) on the test dataset.
   Not very great on the test score, even though the prediction plot looks pretty
   

2. MLP with increased window, lookback = 3

Given the current time (t) wewant to predict the value at the next time in the sequence (t+1), we can use the current time(t) as well as the two prior times (t-1 and t-2). Increasd thelookback argument from 1 to 3.
Train Score: 21.29 RMSE Test Score: 45.04 RMSE
The model has an average error of 21 passengers (in thousands) on the training dataset and 45 passengers (inthousands) on the test dataset. Improved.

3. LSTM with lookback = 1

Prepare the data Train Score: 22.68 RMSE Test Score: 50.51 RMSE Similar to simple MLP

4. LSTM with larger window lookback = 3

Train Score: 21.12 RMSE Test Score: 61.21 RMSE Worse performance

5. LSTM with time steps

Train Score: 25.76 RMSE Test Score: 57.09 RMSE Worse performance

6. LSTM with memory batches between, stateful = True

Train Score: 21.40 RMSE Test Score: 50.11 RMSE Not better

7. stacked LSTM

Train Score: 26.54 RMSE Test Score: 159.63 RMSE Much Worse on the test score.

Conclusion: MLP with a larger lookback window makes a better prediction

With the final error of 21 passengers (in thousands) on the training dataset and 45 passengers (in thousands) on the test dataset.

Future work

Increase epochs, and tune hyperparameters to increase performance