aimas-upb / data4water

data4water

data4water's Issues

Create plot support

Why

We need to have exploratory plot support in order to visualize the results of our analyses.

What

Implement support for at least the following types of graph:

• boxplot view per hour, where the data for each box are the water consumption values for that given hour for all recorded days
• concatenated view of all water consumption for a given costumer (i.e. concatenate all rows in one single large timeseries - useful for our seasonality analysis)
• plots for timeseries decomposition - view trend, seasonality and residual components

Notes

To make visualization interactive, we will be using the plotly library (install via conda here).

Predict consumption levels within costumer cluster

Why

For a cluster of customer resulting from #3 we want to train models to predict the following 1, 2 or more hours of consumption, based on the intra-day history.

What

• choose prediction lookahead window (1h, 2h, etc.)
• compare results between a classic approach (ARIMA, etc.) and RNN based ones

Investigate seasonality in recorded timeseries

Why

We assume that within the 3 months worth of recorded data there will be shifts in consumption behavior of a costumer.
We want to detect such shifts and extract the corresponding consumption periods to perform additional analysis over each period type.

What

Attempt 3 approaches:

• Seasonality analysis (CT)
  • concatenate all time series for each customer
  • perform decomposition time series
  • extract seasonal windows
  • investigate seasonal component filtering options
• Investigate time series periodicity [MT]
  • start from ideas described here
• Investigate periodicity by clustering (AS)

Anomaly detection within customer clusters

Why

For the customer clusters determined in #3 we want to implement methods to detect abnormal consumption behavior.

What

• TBD

Clustering of customers for each seasonal window

Why

For each period resulting from the analysis in #2 we want to determine all customers who have similar consumption patterns.

What

• implement method to cluster costumers within a given consumption period
  • determine clustering algorithm
  • fine tune algorithm meta-parameters
  • compare outcomes