aivcec / TimeSeriesForecast

Time Series Forecasting in Sensor Data Streams

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Sensory Data Time Series Forecast

An excerpt from my thesis where I was tasked to create a machine learning based forecast for water level and flow of river Kupa in Croatia.

The case study was about using upstream data about water level, flow and raindrop to create a forecast for river data downstream. The multivariate forecast was done using the statistical VAR model and the LSTM model.

Data


The time series data was taken from 4 hydrological stations and various meteorological stations along the stream of river Kupa. The received data consisted of daily readings of 6 years of rainfall(mm), flow(m^3/2) and water level(cm). The first 5 years of data were used as training for the models and the last year is used as test data in these samples.

Files


The forecasting methods are presented in a couple of python notebooks.

  • Series-Analysis.ipynb - shows the methods to infer the stationarity of the series
  • Baseline.ipynb - shows the baseline models used for the forecast
  • VAR-one-step.ipynb - contains the one step VAR model
  • LSTM-one-step.ipynb - contains the one step LSTM model

Also, couple of other .py files are used to hold utility methods.

Dependencies


  • numpy
  • pandas
  • tensorflow 1.1.0
  • keras 2.1.5
  • statsmodels 0.9.0

References


https://www.otexts.org/fpp

https://machinelearningmastery.com/introduction-to-time-series-forecasting-with-python/

About

Time Series Forecasting in Sensor Data Streams


Languages

Language:Jupyter Notebook 99.7%Language:Python 0.3%